Deep learning cross-applicability of Uncrewed Aircraft System (UAS) imagery from different disaster types for building damage assessment

Deep learning cross-applicability of Uncrewed Aircraft System (UAS) imagery from different disaster types for building damage assessment | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep learning cross-applicability of Uncrewed Aircraft System (UAS) imagery from different disaster types for building damage assessment Dae Kun Kang, Michael J. Olsen, Erica Fischer, Jaehoon Jung, Julie A. Adams This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6567339/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 15 You are reading this latest preprint version Abstract In recent years, the frequency and intensity of natural disasters have increased worldwide. These disasters cause significant economic losses, with building damage accounting for a substantial portion. Post disaster response plans involve inventories of building damage. However, these assessments can be highly subjective, require substantial time, and can expose the inspectors to unsafe environments. Automation of building damage assessment by applying deep learning combined with advanced remote sensing technology is currently an active research topic. These efforts are hindered by the limited amount of high-quality training datasets for each disaster type (e.g., hurricane, wildfire). Buildings damaged by different disasters may show distinct damage patterns due to differing damage mechanisms, posing challenges to data integration across disaster types and model development. To investigate these issues, this study explores the interrelationship between wildfire and hurricane data by developing models suited to wildfire and hurricane datasets both individually and jointly as well as combining various backbones and deep learning models. Our approach includes semantic segmentation for pixel-level damage assessment and analyzing model sensitivity with different amounts of training data. Ultimately, this study provides a solution to the limited data available to train building damage assessment deep learning models by providing a comparative analysis of the inter-applicability of wildfire and hurricane data. A notable finding is that when using a small portion of data through transfer learning, data and deep learning models from the other disaster types can be leveraged. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Introduction and background Natural disasters, including wildfires, hurricanes, tornadoes, earthquakes, and floods, are prevalent globally, and increasing in frequency [ 1 ]. For example, in the United States, 28 weather and climate disasters with losses exceeding USD 1 billion dollars in 2023 breaking the previous record of 22 set in 2020 [ 2 ]. These disasters incurred a minimum cost of USD 92.9 billion [ 2 ]. Building damage accounts for a large portion of these costs. Identifying damage to buildings provides one of the costs associated with the disaster and also provides the basis for identifying affected population and determining an appropriate response [ 3 ]. However, the collation of disaster-related information in the field is time-consuming and labor-intensive and often subjective [ 4 ]. Remote sensing technology from satellites and aircraft has provided timely, visual data about disaster areas, to assist with damage assessment. While more efficient and less subjective compared with manual methods, labeling and analyses of these data are time consuming. By analyzing the visual elements embedded in these data, computer vision technologies such as deep learning have come to play a pivotal role in addressing these shortcomings in post-disaster building damage assessment, especially when combined with rapidly advancing remote sensing technologies [ 5 ]. Large databases of labeled images are necessary to support these efforts. For example, the xBD dataset is a building damage assessment dataset containing pre- and post-event satellite images of disaster-related events (earthquake, tsunami, flood, wildfire, hurricane, and volcanic eruption) [ 6 ]. To date, apart from the xBD dataset, there is no other integrated dataset that encompasses a wide range of disasters for post disaster building damage assessment. Although, there are some event specific datasets available from post reconnaissance activities [ 7 , 8 ], most of these have not been labeled to support deep learning analyses. Current research has focused on post disaster building assessment with deep learning and remote sensing data concentrates on singular disaster types rather than looking at disasters holistically. For example, in exploring wildfire damage, Galanis et al. (2021) used a subset of the xBD dataset to apply ResNet (residual network), a convolutional neural network (CNN), to satellite images to classify buildings as damaged and undamaged [ 9 ]. As another example, Cao and Choe (2020) applied a fine-tuned Visual Graphics Group (VGG)-16 to satellite images containing hurricane-damaged buildings and classified the buildings into damaged and undamaged categories [ 10 ]. These studies showed high accuracy in classifying buildings damaged by each disaster, 93% in wildfire and 97% in hurricane. However, they did not provide pixel-level classification, which could provide more detailed information about the disaster area. Additionally, their performance was not verified on high-resolution images because they trained and tested based on satellite images. Moradi and Shah-Hosseini (2020) classified earthquake-damaged buildings into damaged and undamaged buildings using Very High Resolution (VHR) satellite images of Haiti and UNet [ 11 ]. They provided pixel-level classification but used images before and after the earthquake simultaneously to improve accuracy. This has limitations in that high resolution images before the disaster are not always available and the image processing time is longer. Rahnemoonfar et al. (2021) classified flooded buildings using deep learning through a pixel-level analysis using several semantic segmentation modules (PSPNet, Enet, and DeepLabv3 plus) [ 12 ]. Pi et al. (2021) also provided pixel-wise analysis of flooded buildings in flood scenes but utilized video footage rather than static images [ 13 ]. Although most previous studies address a single disaster, they provide the groundwork to advance their models and methodologies to be applicable to a variety of disasters as a future research direction. Recent studies have tried to integrate data across multiple disasters during model training to improve generalization capability. Deng et al. (2022) trained a deep learning model using data from multiple disasters in the xBD satellite dataset simultaneously [ 14 ]. They divided the building damage classification into two stages building segmentation using pre-disaster images with UNet, and classification using pre- and post-disaster images. This method showed high accuracy of 87% in building localization and 75% in classification. Irwansyah et al. (2023) used the xBD dataset and performed building damage assessment through two steps: building localization with UNet and PSPNet and classification with ResNet50 [ 15 ]. However, since they compared the accuracy using only one backbone for UNet and PSPNet, there is potential for improvement by applying several backbones. Xia et al. (2023) proposed a BDANet model that integrates UNet and Cross-directional attention (CDA) module structure [ 16 ]. Although they did not apply their model to several disasters, they used the xBD dataset to train the model with images from multiple disasters. They achieved an average accuracy of 80% when applying the model to post-earthquake images from the Islahiye region of Turkey. Several limitations arise in these prior works. First, many require both pre- and post-disaster images, which may not be available. Second, the classification task is performed in two stages, requiring substantial computation time. While semantic segmentation allows the original image to be used without preprocessing, the two-step approach fails to capitalize on this benefit. For example, extracting buildings as a preprocessing step undermines the efficiency that semantic segmentation offers. Third, models trained with xBD may have potential difficulties in scaling to high-resolution imagery given that xBD contains lower resolution satellite imagery that cannot capture details on damaged structures or debris piles that are important to the classification process. Another key challenge in the field of disaster damage assessment using deep learning is developing models that can be generalized and applied to multiple disaster types. However, there has been relatively little research on the mutual applicability of different disaster data types, which is fundamental for solving this problem. Investigating the mutual applicability of several types of disaster data can provide a basis for alleviating data scarcity, a fundamental issue in post-disaster building damage assessment using deep learning. By understanding the interrelationships between different disaster datasets, more effective data utilization can be achieved. In previous studies such as Irwansyah et al. (2023) [ 15 ], a vast amount of data from multiple disasters were simultaneously used for training; hence, it is difficult to identify the contribution of each disaster type. On the one hand, the use of additional images from other disaster types may substantially increase computational time with limited benefit to accuracy. Also, it is possible that the data from one disaster type may be detrimental to the classification of building damage from another disaster type given the different mechanisms. When categorizing data solely based on the damage extent the different damage mechanisms from each disaster type are neglected in ‘the training stage’. For instance, in a wildfire or tornado, most of the destroyed buildings will no longer be present at the site itself, with just the foundation and some debris remaining (Fig. 4.1a). In contrast, in a hurricane, destroyed buildings may have substantial roof damage but remain mostly intact or consist of a large debris pile (Fig. 4.1b). (a) Wildfire affected houses (b) Hurricane affected houses Figure 4.1 Houses damaged by Wildfire and Hurricane These observations raise a critical question central to this research: Are the differences in damage mechanisms and visual patterns across disaster types substantial enough to hinder the effective cross-application of data? In other words, despite the potential benefits of alleviating data scarcity, can data from one disaster type truly support damage assessment in another, or do these inherent disparities ultimately limit mutual applicability? 2.1 Research Objectives In this study, high-resolution images acquired from UAS and semantic segmentation models were used to evaluate building damage at the pixel level for efficient and accurate damage assessment in complex disaster environments. In addition, through two datasets, wildfire and hurricane, we investigated the mutual applicability between disasters in a deep learning environment from two perspectives: the dataset and the model. The objectives of this study are (1) investigate the interrelationship between wildfire and hurricane data to explore the influence of different disaster data on analyzing buildings damaged by each disaster; (2) investigate the impact of the type and combination of backbones and models on accuracy; and (3) evaluate the inter-applicability of deep learning models developed from wildfire and hurricane data through transfer learning to extend a model developed from one disaster type to another. In this process, the effectiveness of the model relative to the amount of additional training data used is analyzed. Ultimately, this study contributes by providing a framework to demonstrate how to expand the data available for post-disaster building damage assessment by integrating different disaster datasets and can be extended to other disaster types. Methods This study consists of four major parts (Fig. 4.2 ). First, in the data preparation stage, the wildfire and hurricane datasets to be used in this study were prepared with pixel-level annotations according to the degree of damage to buildings (standing and destroyed) and the characteristics of building damage from each disaster. Then, the backbones and deep learning models, loss function, accuracy assessment method, and the model training environment used to analyze these datasets were prepared. Second, several combinations of backbones and models were tested to find a deep learning model that is well suited for both wildfire and hurricane datasets. Third, the selected deep learning model was trained using the wildfire dataset, hurricane dataset, and the wildfire and hurricane integrated dataset. This model was then applied to each disaster dataset to analyze the impact of wildfire and hurricane data on each other in the deep learning environment. Fourth, the model developed from one disaster was applied to another disaster using transfer learning. The amount of training data used was varied and the associated sensitivity to the amount of training data was analyzed. 3.1 Datasets To illustrate the framework, we utilize high-resolution Uncrewed Aircraft System (UAS) datasets obtained from two disaster types: wildfire and hurricane. These datasets consist of UAS images and masks containing pixel-level annotations of buildings for semantic segmentation annotation. Buildings are classified into standing and destroyed. 3.1.1 Wildfire dataset Wildfire images were obtained from the Marshall fire, which burned over 6,080 acres and destroyed 1,056 structures in Boulder County, Colorado on December 30, 2021 [ 17 , 18 ]. Two UAS systems were utilized. First, a DJI Matrice 210 V2 RTK with X5S camera, a medium-sized UAS, captured high resolution imagery for a small town close to Mulberry St. Second, a Sensefly eBee X, a larger UAS with a flight endurance of up to one hour, was used to map large areas across the study area. The Roboflow platform was employed for the annotation of dataset for semantic segmentation [ 19 ]. All buildings were manually digitized as a polygon along the building boundary. Three expert annotators were introduced, and the annotations were checked and revised by the authors before final approval. FEMA guidelines classify building damage into four levels: Affected, Minor, Major, and Destroyed. This generalized classification encompasses all disasters other than floods. However, most buildings in areas affected by wildfire generally are either completely burned down or show minimal damage. As a result, we classified the building damage into two categories. Standing, which referred to buildings with no observed damage or buildings requiring minimal repairs to make them habitable and destroyed, which referred to buildings requiring extensive repairs or buildings where the structural framing of the roof or walls has been damaged, leaving the interior exposed. Figure 4.3 shows the image and mask pairs of wildfire damaged buildings. Green represents standing buildings and red represents destroyed buildings. 3.1.2 Hurricane dataset RescueNet [ 5 ], contains detailed imagery captured following Hurricane Michael by the Center for Robot-Assisted Search and Rescue using small UAS (DJI Mavic Pro quadcopters) at Mexico Beach and other directly impacted areas. A total of 80 flights were performed between October 11–14, 2018. Hurricane Michael made landfall on October 10, 2018, near Panama City, Florida, with sustained winds of approximately 160 mph. This event caused upwards of 25 billion dollars in damages with over 54,000 structures damaged or destroyed [ 20 ]. The original RescueNet dataset classified buildings with hurricane damage into four levels: Building-No-Damage, Building-Medium-Damage, Building-Major-Damage, and Building-Total-Destruction. However, to directly compare with the wildfire datasets in this study the categories of Building-No-Damage and Building-Medium-Damage were integrated into Standing and Building-Major-Damage and Building-Total-Destruction were integrated into Destroyed (Fig. 4.4 ). 3.2 Preparation Figure 4.5 provides an overview of the dataset preparation process for this study. In total, we prepared 3,789 wildfire images and 4,494 hurricane images with pixel-level annotations for buildings. Images were patched into 256x256 segments. These patches were divided 70:15:15 for deep learning model training, validation, and testing. Note that to maintain balance and avoid overfitting to a specific class, the division maintained a consistent ratio of standing buildings and destroyed buildings in each subset both in context of the number of polygons and pixels representing structures. Next, data augmentation was applied to these data through nondestructive transformations (e.g., rotate 90, 180, and 270 degrees, horizontal, vertical, and diagonal reflection) to maintain the information of the original image and mask. In this study, the wildfire dataset is represented as W dataset, and the wildfire test dataset used for model performance validation is represented as W test dataset. The hurricane dataset is also represented similarly: H dataset and H test dataset. 3.3 Model training 3.3.1 Semantic segmentation network architectures U-Net-based architectures have been used and proven in the field of post-disaster building damage assessment [ 5 , 14 , 15 , 21 , 22 , 23 , 24 , 25 ]. The U-Net, Attention U-Net, and U-Net 3 plus were explored for the base model (Table 4.1 ). U-Net is a convolutional neural network designed for image segmentation, developed by Ronneberger et al. in 2015 [ 26 ]. Its U-shaped architecture includes a contracting path (encoder) and an expansive path (decoder). The U-Net architecture is distinguished by its ability to extract image features and enable precise localization simultaneously. This is achieved through the integration of both low-dimensional and high-dimensional information. A key mechanism in U-Net is the use of skip connections, where features from each layer in the encoder are directly merged with the corresponding layer in the decoder. This concatenation method enhances the model’s capacity to retain spatial information and improve segmentation accuracy. U-Net's skip connections and precise localization capabilities allow it to accurately identify both the shapes of the entire building and the details of the damage. Attention U-Net enhances the standard U-Net architecture by incorporating attention mechanisms the network to focus on relevant parts of the input image, thereby improving segmentation performance, particularly in cases involving small or complex structures [ 27 ]. The architecture introduces attention gates (AGs) into the standard U-Net design. AGs are employed during the skip connections procedure to focus on relevant features from the encoder before merging them with the decoder's features. These take features from both the encoder and the decoder, using them to generate an attention map that highlights important regions while suppressing irrelevant ones. The attended features are then concatenated with the upsampled features from the decoder. This improved focus is expected to help the network better capture contextual information and segment complex damaged buildings more effectively. U-Net 3 plus builds upon the U-Net and U-Net 2 plus architectures by introducing full-scale skip connections, aiming to make full use of multi-scale features for more accurate segmentation [ 28 ]. U-Net 3 plus redesigns the interconnections between the encoder and decoder, as well as the interconnections between decoder sub-networks. This enables the capture of fine-grained details and coarse-grained semantics across all scales. Each decoder layer incorporates feature maps from smaller- and same-scale encoder layers, and larger-scale decoder layers, merging them to enhance the segmentation performance. This approach is expected to be effective for buildings appearing at varying scales. However, the accuracy of the models may vary depending on the combination of various backbones. A backbone refers to the established architecture or network employed for feature extraction, having been trained on numerous tasks previously and proven its efficacy [ 29 ]. The following backbone networks were used in this study: Residual Neural Network (ResNet), Densely Connected Convolutional Network (DenseNet), and EfficientNet (Table 4.1 ). ResNet is a model based on convolutional neural networks (CNNs), featuring the introduction of residual networks [ 30 ]. ResNet has various versions depending on the number of parameter layers. ResNet includes skip connections or recurrent units between blocks of convolutional and pooling layers. Additionally, each block is followed by batch normalization [ 31 ]. ResNet-101 with 44.7M parameters and ResNet-152 with 60.4M parameters are adopted in this study. DenseNet is CNN architecture designed to enhance the flow of information and gradients throughout the network by creating direct connections between all layers [ 32 ]. This facilitates the training of deeper and more complex networks. In DenseNet, each layer takes inputs from all previous layers, promoting feature reuse and improving gradient propagation, which leads to more efficient and effective training. DenseNet models tend to be more compact and require fewer parameters compared to traditional CNNs, making them ideal for deployment in environments with limited resources. DenseNet-169 with 14.3M parameters and DenseNet-201 with 20.2M parameters are adopted in this study. EfficientNet is a modern family of architectures, showing outstanding performance in classification tasks while utilizing fewer parameters and Floating Point Operation (FLOP) than other networks [ 33 ]. It employs compound scaling to efficiently and uniformly adjust the network's width, depth, and resolution. The various versions of EfficientNet, from B0 to B7, can be chosen based on resource availability and computational cost. EfficientNet-B4 with 19.5M parameters and EfficientNet-B7 with 66.7M parameters are adopted in this study. In this study, we experimentally developed models that is well-applied to wildfire and hurricane UAS datasets through the combination of three types of U-Net architectures (U-Net, Attention U-Net, and U-Net 3 plus) and six backbones (ResNet-101, ResNet-152, DenseNet-169, DenseNet-201, EfficientNet-B4, and EfficientNet-B7). 3.3.2 Loss function The loss represents the total errors from each batch in the training or validation sets, indicating how well or poorly the trained model performs following each optimization step [ 34 ]. In choosing the loss function, we focused on data imbalance both between the pixels of the standing building and the destroyed building and between these categories and the background pixels. We used the focal loss function (Categorical Focal Cross entropy) proposed by Lin et al. (2017) where the number of pixels are weighted inversely, thereby reducing the influence of the background, which occupies a large portion of each image [ 35 ]. Table 4.1 Summary of deep learning models and backbone networks used in the study Category Architectures Details Deep Learning Models U-Net [ 26 ] - CNN for image segmentation with U-shaped architecture. - Contracting (encoder) and expansive (decoder) paths. - Skip connections for precise localization. Attention U-Net [ 27 ] - Enhanced U-Net with attention mechanisms. - Focus on relevant image parts for better segmentation. - Improved performance on complex or small structures. U-Net 3 Plus [ 28 ] - Build on U-Net and U-Net 2 Plus. - Full-scale skip connections. - Capture fine-grained details and coarse semantics across scales. Backbone Networks ResNet [ 30 ] - CNN with residual networks. - Skip connections for efficient feature extraction. DenseNet [ 32 ] - CNN architecture with direct connections between all layers. - Compact design ideal for resource-limited environments. EfficientNet [ 33 ] - Modern architecture family with outstanding classification performance. - Utilize compound scaling to adjust network width, depth, and resolution. - Fewer parameters and FLOPs. 3.3.3 Accuracy assessment To evaluate the performance of our model, the following four metrics (accuracy, precision, recall, and F1 score) were calculated at the pixel level for each class (standing and destroyed) using Equations 4.1–4.4 [ 36 ]. These factors consider the number of True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN). \(\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\:\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}\) Eq 4.1 \(\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}=\:\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{N}}\) Eq 4.2 \(\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\:\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{F}\text{N}+\text{F}\text{P}+\text{T}\text{N}}\) Eq 4.3 \(\:\text{F}1=\frac{2\text{*}\text{T}\text{P}}{2\text{*}\text{T}\text{P}+\text{F}\text{P}+\text{F}\text{N}}\) Eq 4.4 Precision (Eq. 4.1) describes how many standing or destroyed building pixels are detected as standing or destroyed. Recall (Eq. 4.2) describes how many of the actual standing or destroyed building pixels have been detected. Accuracy (Eq. 4.3) describes how many pixels have been correctly detected, regardless of whether they are destroyed or standing. The F1 score (Eq. 4.4) provides a measure of the harmonic mean of precision and recall, which can reduce bias resulting from data imbalance. Hence, the F1 score was the primary factor considered to determine the model to reduce the impact of the imbalance in disaster datasets. 3.3.4 Setup for model training Keras with the TensorFlow was employed as a framework with the adaptive moment estimation (Adam) optimization algorithm [ 37 ]. During the 100 epochs of training, validation data was sent across the network, and its loss was estimated and monitored. The network was trained in 16 batches until it reached convergence at an initial learning rate of 1e-4. To avoid overfitting, training was stopped when the validation loss did not increase for 5 epochs and saved the model with the smallest validation loss. Epochs, learning rate, and batch size were experimentally determined through repeated testing. Because the datasets used in this study were very large, limitations in computing power and time were important factors in determining these variables. 3.4 Base model and backbone evaluation Figure 4.6 shows the overall process to combine the different backbones and deep learning models to determine the most suitable model for each disaster. A total of 18 models, combining 6 backbones and 3 deep learning models, were applied to each disaster dataset prepared through the dataset preparation process. Through these accuracy assessment metrics (accuracy, precision, recall, and F1 score) computed for each model, the overall performance for each disaster was evaluated. To select the optimal combination of models for each disaster, each was ranked by its F1 score, and the model with the lowest sum of ranks was selected as the base model. This process results in a total of three model combinations: 1) The optimal W model: the model that performed best on the W dataset, 2) the optimal H model: the model that performed best on the H dataset, and 3) the base model: the model that performed well overall on both disasters. The base model was used to analyze the cross-applicability of different types of disaster datasets. The optimal W and H models were subsequently used for cross-applicability of different types of disaster models with transfer learning. 3.5 Cross-applicability of different types of disaster datasets This section describes the method of investigating the interrelation between different disaster datasets in a deep learning environment as a fundamental study for the integrated use of different disaster data. By analyzing results using different disaster datasets for training both separately and together, we can examine how each type of disaster data influences the analysis of the other. Figure 4.7 shows the overall process for investigating cross-applicability of different disaster types. Each dataset was trained with the base model determined in the previous section. The three trained models (wildfire model (W model), hurricane mode (H model), and combined model (W + H model)) were then applied to the W and H test datasets to evaluate their performance by changing the training dataset while using the same base model. 3.6 Transfer learning While the previous section investigated the mutual applicability between disaster datasets, this section focuses on whether the use of transfer learning is a more efficient and effective method to apply a model developed from one disaster to another type of disaster. In this section, we cross-applied the optimal W and H models to each other to investigate the accuracy according to whether transfer learning is applied and the amount of training data used in order to apply a model developed from one disaster to another disaster (Fig. 4.8 ). To apply the optimal H model to the W test dataset, we re-trained the optimal H model with 1%, 3%, 5%, 10%, 20%, 30%, 50%, 75%, 100% of the entire W training dataset. We then analyzed the sensitivity of the model according to the data size. This same experiment was conducted both with and without transfer learning. To apply the optimal W model to the H test dataset, this same procedure was followed. Results and Discussion 4.1 Base model decision Table 4.2 shows the statistics (precision, recall, accuracy, and F1-score) for all 18 combinations of the backbones and deep learning models on H and W datasets. From the results in Table 4.2 , we were able to find notable trends. The accuracy difference according to the number of backbone parameter layers was not significant. Although the difference in the number of parameters between the two versions was 6M for DenseNet at the minimum and 47.2M for EfficientNet at the maximum, differences in the F1 score were not significant overall, ranging from 0.0003 to 0.0262. Notably, the higher version with more parameters did not always show higher accuracy. The accuracy difference according to the backbone was larger than the accuracy difference according to the model. The difference in F1 score of the three models using the same backbone was from 0.0067 to 0.0201 on the H dataset, and from 0.0067 to 0.0290 on the W dataset. On the other hand, the differences in F1 scores of the six variations of the same model ranged from 0.0412 to 0.0650 on the H dataset and from 0.0390 to 0.0999 on the W dataset. This suggests that modifying the backbone of an existing model in post-disaster building damage assessment can potentially enhance accuracy. The model accuracy was average 0.0846 higher on the W dataset than on the H dataset. This issue may arise because the damage patterns in the H dataset are more complex and diverse than those in the W dataset. For example, it is relatively easy to identify the structural boundaries of buildings damaged by wildfires compared to buildings damaged by hurricanes, which results in differences in the accuracy of building annotations. Table 4.2 Evaluation metrics of all combinations of the backbones and deep learning models on hurricane and wildfire datasets (Important numbers are in bold) Model Backbone Precision Recall Accuarcy F1-Socre Hurricane Wildfire Hurricane Wildfire Hurricane Wildfire Hurricane Wildfire U-Net ResNet 101 0.8204 0.9104 0.7685 0.8304 0.9664 0.9869 0.7934 0.8671 ResNet 152 0.8215 0.8955 0.7759 0.8887 0.9669 0.9884 0.7980 0.8921 DenseNet 169 0.7838 0.8905 0.7748 0.8483 0.9622 0.9862 0.7785 0.8686 DenseNet 201 0.7614 0.8782 0.8013 0.8617 0.9610 0.9858 0.7789 0.8698 EfficientNet B4 0.8160 0.9100 0.8165 0.8846 0.9688 0.9893 0.8162 0.8971 EfficientNet B7 0.8269 0.8998 0.8127 0.9129 0.9698 0.9899 0.8197 0.9062 Attention U-Net ResNet 101 0.8275 0.8968 0.7756 0.8792 0.9676 0.9879 0.8000 0.8879 ResNet 152 0.8244 0.8715 0.7831 0.8877 0.9678 0.9866 0.8024 0.8795 DenseNet 169 0.8058 0.8810 0.7436 0.8443 0.9633 0.9849 0.7726 0.8622 DenseNet 201 0.7944 0.8764 0.7381 0.7383 0.9619 0.9781 0.7651 0.8014 EfficientNet B4 0.8386 0.9114 0.8077 0.8913 0.9706 0.9896 0.8229 0.9013 EfficientNet B7 0.8443 0.9183 0.7855 0.8796 0.9700 0.9896 0.8135 0.8984 U-Net 3 plus ResNet 101 0.7333 0.9031 0.8363 0.8893 0.9596 0.9889 0.7799 0.8961 ResNet 152 0.8430 0.9005 0.7742 0.8876 0.9691 0.9887 0.8062 0.8940 DenseNet 169 0.8041 0.8808 0.7625 0.8493 0.9641 0.9849 0.7827 0.8648 DenseNet 201 0.8185 0.8807 0.7216 0.8643 0.9635 0.9861 0.7659 0.8724 EfficientNet B4 0.8570 0.9158 0.7875 0.8961 0.9712 0.9902 0.8201 0.9058 EfficientNet B7 0.8420 0.9108 0.8200 0.9004 0.9716 0.9901 0.8308 0.9056 Table 4.3 shows the ranking of the combinations of backbones and deep learning models based on F1 score. As a result, U-Net 3 plus with EfficientNet B7 showed the highest performance with 0.8308 in the H dataset (the optimal H model), and U-Net with EfficientNet B7 showed the highest performance with 0.9062 in the W dataset (the optimal W model). Based on the sum of the ranks, U-Net 3 plus with EfficientNet B7 was selected as the base model because it performed well overall on both datasets. Table 4.3 Ranking of the combinations of backbones and deep learning models based on F1 score (Important numbers are in bold) Model Backbone Hurricane Rank Wildfire Rank Sum Total Rank U-Net ResNet 101 0.7934 11 0.8671 15 26 12 ResNet 152 0.7980 10 0.8921 9 19 8 DenseNet 169 0.7785 15 0.8686 14 29 15 DenseNet 201 0.7789 14 0.8698 13 27 13 EfficientNet B4 0.8162 5 0.8971 6 11 5 EfficientNet B7 0.8197 4 0.9062 1 5 2 Attention U-Net ResNet 101 0.8000 9 0.8879 10 19 8 ResNet 152 0.8024 8 0.8795 11 19 8 DenseNet 169 0.7726 16 0.8622 17 33 17 DenseNet 201 0.7651 18 0.8014 18 36 18 EfficientNet B4 0.8229 2 0.9013 4 6 4 EfficientNet B7 0.8135 6 0.8984 5 11 5 U-Net 3 plus ResNet 101 0.7799 13 0.8961 7 20 11 ResNet 152 0.8062 7 0.8940 8 15 7 DenseNet 169 0.7827 12 0.8648 16 28 14 DenseNet 201 0.7659 17 0.8724 12 29 15 EfficientNet B4 0.8201 3 0.9058 2 5 2 EfficientNet B7 0.8308 1 0.9056 3 4 1 4.2 Cross-applicability of different types of disaster datasets 4.2.1 Application to the wildfire dataset Three types of trained models from the base model were developed: the W model, the H model, and W + H model. Table 4.4 and Fig. 4.9 show the results of applying all three models to the W test dataset. The W model achieved high performance on the wildfire test dataset, confirmed by an F1 score of 0.9527 for the standing building class, 0.8507 for the destroyed building class, and 0.9017 on average. Also, comparing the ground truth in Fig. 4.9 b and the result of applying the W model in Fig. 4.9 c, we can visually confirm that this model detects both standing and destroyed buildings well. However, the H model showed very poor performance on the W test dataset (average F1 = 0.2903), particularly for destroyed buildings (F1 = 0.0153). As Fig. 4.9 d shows, the H model hardly detects buildings damaged by wildfire. However, the W + H model, trained on a combined dataset of W and H data, achieved similarly high performance to the W model on the W test dataset (average F1 = 0.8936). Table 4.4 Evaluation metrics for the wildfire model (W), hurricane model (H), and combined model (W + H) on the wildfire test dataset Model Class Precision Recall Accuarcy F1-Score W model Standing 0.9604 0.9452 0.9925 0.9527 Destroyed 0.8537 0.8476 0.9868 0.8507 Average 0.9071 0.8964 0.9902 0.9017 H model Standing 0.7708 0.4463 0.9453 0.5653 Destroyed 0.0354 0.0097 0.9445 0.0153 Average 0.4031 0.2280 0.9396 0.2903 H + W model Standing 0.9478 0.9451 0.9915 0.9464 Destroyed 0.8559 0.8263 0.9862 0.8408 Average 0.9018 0.8857 0.9886 0.8936 4.2.2 Application to the hurricane dataset Table 4.5 and Fig. 4.10 present the results of applying the H model, the W model, and the W + H model to the H test dataset. The overall trends were similar to when the models were applied to the W test dataset. The H model achieved high performance with an average F1 score of 0.8261 on the H test dataset. The W model showed low performance with an F1 score of 0.4302, which is similar to when the H model was cross applied to the W test dataset. Nevertheless, the H + W model achieved a similar performance to the H model with an F1 score of 0.8268 on the H test dataset. Table 4.5 Evaluation metrics for the hurricane model (H), wildfire model (W), and combined model (W + H) on the hurricane test dataset Model Class Precision Recall Accuarcy F1-Score H model Standing 0.9278 0.8697 0.9747 0.8978 Destroyed 0.7662 0.7428 0.9676 0.7543 Average 0.8470 0.8063 0.9712 0.8261 W model Standing 0.6967 0.7486 0.9264 0.7217 Destroyed 0.1661 0.1191 0.9008 0.1388 Average 0.4314 0.4339 0.9136 0.4302 H + W model Standing 0.9018 0.9011 0.9749 0.9015 Destroyed 0.7907 0.7172 0.9683 0.7522 Average 0.8463 0.8092 0.9716 0.8268 4.2.3 Comparison The H model, the W model, and the W + H model were applied to the W and H test datasets, respectively, and the performances were statistically and visually analyzed. The results experimentally demonstrated that the damage patterns of buildings affected by hurricanes and wildfires are distinguished and cannot be ignored in a deep learning environment. Due to the complexity of deep neural networks, pinpointing the exact factors behind these results was challenging [ 38 ]. Therefore, in this study, we conducted a comparison analysis using the predicted result images to investigate the causes. This approach allowed us to identify two notable characteristics. The first is the difference in the characteristics of the debris. Debris of buildings damaged by wildfire is dominated by a damage pattern in which they are chemically transformed due to the influence of intense heat. However, in the case of hurricanes, physical transformation is dominant. As Fig. 4.11 shows, the debris of buildings damaged by hurricanes and by wildfires differ in shape and size, which may have made it difficult for the H model trained on hurricane data to detect the debris of buildings damaged by wildfires. The second factor is the distribution of debris. Most of the debris from buildings damaged by wildfire remain within the building boundary. In contrast, hurricane-induced debris often extends beyond the building boundary due to strong winds. This difference in debris distribution patterns can negatively affect the accuracy when the W model is applied to H data, as it may misinterpret debris beyond the building boundary as a destroyed building. This can be visually confirmed through the difference between Fig. 4.12a and b. Figure 4.12 Example results of the W model applied to the wildfire and hurricane test datasets highlighting the differences in debris distribution. Despite these differences, the W + H model achieved high performance on both disaster datasets. Paradoxically, these results highlight the importance of investigating the interrelationships between datasets when integrating diverse disaster data. Combining these datasets without understanding their interrelationships can lead to an increase in data volume that does not improve accuracy and, consequently, prolongs model training time. Notably, training the model using the combined dataset took twice as long as training the model using a single dataset. 4.3 Transfer learning In the previous section, we analyzed the impact of cross-applying W and H datasets from the perspective of the datasets, while in this section, we analyzed it from the perspective of deep learning models. We investigate how to cross-apply the optimal single disaster models to each other, so that a model developed from one disaster can be used for another. 4.3.1 Application to the wildfire dataset Figure 4.13 shows the accuracy (F1 score) of the optimal H model applied to the W test dataset. However, unlike the previous section where the H model was directly applied to the W test dataset, the optimal H model was trained using the pre-trained weights on the H dataset using transfer learning. Portions of the W dataset ranging from 1–100% were utilized in the transfer learning process. Figure 13a and b shows an experiment with and another without transfer learning applied to examine its benefits. The experiments revealed several notable results. First, when transfer learning was used, high accuracy could be achieved with nominal additions to the training dataset size. For example, when the model was trained using 133 wildfire images (5%), the F1 score was 0.7504 with transfer learning, which was 0.2669 higher than without transfer learning. Nevertheless, the difference in accuracy becomes smaller as the amount of training data used increases (e.g., difference of 0.1759 for 20%, 0.0348 for 50%, and 0.0031 for 100%). Second, applying transfer learning led to a significant improvement in accuracy with only 1% (27 images) of the training data added. Adding this small amount of training data increased the average F1 score from 0.2835 to 0.6198 (a difference of 0.3363). This improvement was particularly pronounced in the destroyed building class where the F1 score increased from 0.0610 to 0.4793 (a difference of 0.4183). Compared with the highest F1 score (approximately 0.9), 0.6198 might seem low; however, in the early stages of a disaster response, the timeliness of the information can be more important than achieving the highest levels of accuracy. For instance, immediately after a disaster, it may not be necessary to predict damaged buildings with 90% accuracy. If we can assess the damage status of buildings with 60% accuracy across the entire disaster site, this level of accuracy might be sufficient to deploy safety personnel. Understanding that the required accuracy may vary over time, it is a meaningful discovery that we can achieve adequate accuracy by adding a small amount of training data specific to the disaster event as more data are collected and available. Figure 4.13 The F1 score of the optimal H model applied to the W test dataset depending on whether transfer learning was applied and the amount of W training data. 4.3.2 Application to the hurricane dataset Figure 4.14 shows the accuracy (F1 score) of the optimal W model applied to the H test dataset depending on whether transfer learning was applied and with varying the amount of training data. The overall trend of the results of this experiment was similar to when the same experiment with the W dataset. When transfer learning was used, higher accuracy was achieved with the same amount of added training data, but the difference became smaller as the size of the training dataset increased. Similarly, a large increase in accuracy was achieved with a very small addition of training data. When trained using 1% (45 images) of the entire H dataset, the average F1 score improved by 0.1030, from 0.4146 to 0.5176, which is the largest improvement observed. However, this increase is much smaller compared to the improvement in the wildfire case, which was 0.4183. A notable difference is that the slope of the increase in accuracy relative to the amount of training data is more linear and gradual compared with the wildfire dataset. We hypothesize that the reason for these differences result from the fact that the characteristics of hurricane building damage are more diverse than those of wildfire. Figure 4.14 The F1 score of the optimal W model applied to the H test dataset depending on whether transfer learning was applied and the amount of H training data used. 4.3.3 Comparison In the previous section, we conducted experiments to see whether models developed in one disaster can be applied to other disasters using transfer learning. From the results, we were able to find notable trends. First, deep learning models showed higher performance in standing buildings than in destroyed buildings, regardless of whether transfer learning was used. This is because standing buildings are more clearly defined than destroyed buildings. Standing buildings have similar and regular shapes, patterns, and colors, which is advantageous in feature extraction and classification. On the other hand, destroyed buildings show a wide range of variations and irregularities depending on the degree of damage caused by the disaster. As a result, it is difficult to extract common features compared to standing buildings. However, we found that the performance difference between the two categories was large with a small number of training data, but as the number of training data increased, the performance on destroyed buildings improved significantly, approaching the performance on standing buildings. The performance on standing buildings according to the increase in training data was different in hurricanes and wildfires. Comparing Fig. 4.13b and Fig. 4.14b, when using a small number of training data, the performance on standing buildings tends to be higher in hurricanes than in wildfires. This can be seen because the diversity of standing buildings in the dataset is higher in wildfires, which requires more types of data. However, as can be seen in Fig. 4.13a and Fig. 4.14a, it was found that the difference from the diversity of standing buildings, which may vary depending on the disaster or region, can be significantly improved through transfer learning. Finally, the introduction of transfer learning led to a dramatic reduction in training time in both datasets. Table 4.6 shows the training time per epoch of the optimal models applied to the test dataset depending on whether transfer learning was applied and the amount of training data used. When transfer learning was applied, the optimal H model was shortened from 833s to 420s, and the optimal W model was shortened from 1067s to 573s, showing a time reduction of about 2 times in both cases. Also, in both cases, the amount of training data used and the training time per epoch are proportional. Combining this result with the accuracy from the previous section, we can experimentally see how much training data is required to train the model, the training time, and the corresponding accuracy. For example, if we apply the hurricane model to wildfire using transfer learning, it means that with about 1% (27 images) of the training data, we can achieve an accuracy of about 60% (average F1 = 0.6198) and about 660 seconds of training for 100 epochs. The results of these experiments clearly demonstrate the usefulness of transfer learning and its corresponding benefits in the field of disaster assessment. Table 4.6 Training time per epoch of the optimal H and W model applied to the W and H test dataset depending on whether transfer learning was applied and the amount of W and H training data used (TL = with transfer learning; nTL = without transfer learning; H = the optimal H model applied to the W test dataset; W = the optimal W model applied to the H test dataset; Units = sec) Model Method 1% 3% 5% 10% 20% 30% 50% 75% 100% H TL 4 13 21 42 84 126 210 315 420 nTL 8 25 42 83 167 250 417 625 833 W TL 6 17 29 57 115 172 287 430 573 nTL 11 32 53 107 213 320 533 800 1067 4.4 Challenges and future directions One of the primary challenges of this study was the availability of insufficient high-resolution post-disaster datasets that captured building damage. We used datasets obtained from one wildfire (Marshall fire) and one hurricane (Hurricane Michael) event. Therefore, our experimental results may be overfitted to these specific datasets. The simplest solution would be to use datasets from a variety of events, but this is not easily achievable in practice. Efforts to apply deep learning to post-disaster building damage assessment are steadily increasing, especially for those containing high-resolution images [ 5 , 39 ]. With the recent increase in the use of UAS in disaster response, we anticipate that more high-resolution disaster datasets will become available in the future [ 40 ]. However, beyond the challenges in acquiring these data in the short window following a disaster, the annotation process required for deep learning is extremely labor-intensive, which can pose another major barrier. Hence, it is anticipated that annotation of disaster data will remain slow in the near future. In light of these challenges, we made the following efforts were made to minimize overfitting due to insufficient data. First, data augmentation was implemented to increase the diversity of the dataset. Additionally, early stopping was applied during the training process. Lastly, batch normalization was used to normalize the data within each batch based on the mean and variance, ensuring consistency across different distributions during the learning process. Another challenge is the absence of standards for systematic disaster data acquisition and processing. To develop models applicable to multiple types of disasters, the integrated use of diverse disaster data from multiple sources is essential. However, most post-disaster datasets currently acquired using UAS are collected and processed according to the criteria of the respective groups that gather them. These variations in flight pattern strategies, UAS shooting angles and altitudes, and desired resolutions can hinder data integration. While this diversity can eventually be beneficial in increasing data variety when sufficient data has become available, there is currently no benchmark dataset for post-disaster building damage assessment that includes high-resolution images. Establishing standards to facilitate systematic data collection, integration, processing, and labeling is necessary [ 6 , 7 ]. With respect to the annotation process, there is also no common standard for classifying building damage across the different disaster datasets obtained with UAS or other remote sensing data source. Many studies follow FEMA's guidelines, which are based on ground observations and only categorize disasters into flood or others. Since many UAS collections consist of a nadir aerial view, the guidelines and categorization require adaptation. For example, when annotating a building damaged by a hurricane, where should the boundaries of the damaged building be set? Should building damage from all disasters be categorized into four categories or two? Establishing clear guidelines for annotating building damage will ensure consistency and improve the accuracy and reliability of models developed for post-disaster building damage assessment as multiple data sources are combined. Conclusion As foundational research to expand the data availability for post-disaster building damage assessment and to develop a model applicable to various disasters, this study explored the interoperability between different types of high-resolution disaster datasets, focusing on wildfire and hurricane datasets. By evaluating and combining several backbones and deep learning models, we successfully developed models to assess building damage caused by these disasters. In addition, we cross-applied these models to each disaster dataset to explore the possibility of integrating the two disaster datasets. The results experimentally demonstrated that the damage patterns of buildings affected by hurricanes and wildfires are distinct and cannot be ignored in a deep learning environment. Nevertheless, we demonstrated that the models developed from each disaster can be successfully cross-applied through transfer learning. A notable finding was that when cross applying a model developed from one disaster to another, a significant increase in accuracy can be achieved with a very small amount of additional training data when transfer learning is utilized. This accuracy increase was particularly evident on the wildfire dataset. Future studies will focus on expanding the diversity of disaster datasets by including data from several events and sources. Efforts will also be made to standardize the acquisition and processing of high-resolution post-disaster datasets. Furthermore, establishing standard guidelines for annotating building damage across different types of disasters is essential. Additionally, research will continue to explore methods to mitigate overfitting. By addressing these challenges, future work can improve the accuracy and reliability of models for post-disaster building damage assessment, making them applicable to a broader range of disaster scenarios. Declarations Acknowledgements The authors would like to thank the NHERI RAPID Facility (CMMI- 2130997) and GEER team (CMMI-1266418) who helped in data collection and processing. The authors also thank Maryam Rahnemoonfar, Tashnim Chowdhury, and Robin Murphy who shared the hurricane dataset. Also, the authors are grateful to the Federal Aviation Administration (FAA) and Alliance for System Safety of UAS through Research Excellence (ASSURE) for research funding, which indirectly supported this research. Lastly, the authors appreciate the support of Geoffrey Hollinger who provided helpful feedback on the manuscript. Author information Authors and Affiliations School of Civil and Construction Engineering, Oregon State University, Corvallis, OR, USA Dae Kun Kang, Michael J. Olsen, Erica Fischer Department of Urban Engineering, Gyeongsang National University, Jinju, South Korea Jaehoon Jung School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR, USA Julie A. Adams Funding This work was supported by Oregon State University through graduate research assistantships. No other funding was received for conducting this study. Data Availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Ethics approval Not applicable Consent to participate Not applicable Competing interests The authors declare that they have no competing interests to disclose. Consent to publish declaration Not applicable Clinical trial Not applicable References Ritchie, H., Rosado, P. and Roser, M., 2023. Natural disasters. Our World in Data. Smith, A.B. (2024) 2023: A historic year of U.S. billion-dollar weather and climate disasters, NOAA Climate.gov. Available at: https://www.climate.gov/news-features/blogs/beyond-data/2023-historic-year-us-billion-dollar-weather-and-climate-disasters (Accessed: 05 August 2024). Dell'Acqua, F. and Gamba, P., 2012. Remote sensing and earthquake damage assessment: Experiences, limits, and perspectives. Proceedings of the IEEE, 100(10), pp.2876–2890. Lallemant, D., Soden, R., Rubinyi, S., Loos, S., Barns, K. and Bhattacharjee, G., 2017. Post-disaster damage assessments as catalysts for recovery: A look at assessments conducted in the wake of the 2015 Gorkha, Nepal, earthquake. Earthquake Spectra, 33(1_suppl), pp.435–451. Rahnemoonfar, M., Chowdhury, T. and Murphy, R., 2023. RescueNet: a high resolution UAV semantic segmentation dataset for natural disaster damage assessment. Scientific data, 10(1), p.913. Gupta, R., Goodman, B., Patel, N., Hosfelt, R., Sajeev, S., Heim, E., Doshi, J., Lucas, K., Choset, H. and Gaston, M., 2019. Creating xBD: A dataset for assessing building damage from satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (pp. 10–17). Berman, J.W., Wartman, J., Olsen, M., Irish, J.L., Miles, S.B., Tanner, T., Gurley, K., Lowes, L., Bostrom, A., Dafni, J. and Grilliot, M., 2020. Natural hazards reconnaissance with the NHERI RAPID facility. Front Built Environ 6. Wartman, J., Berman, J.W., Bostrom, A., Miles, S., Olsen, M., Gurley, K., Irish, J., Lowes, L., Tanner, T., Dafni, J. and Grilliot, M., 2020. Research needs, challenges, and strategic approaches for natural hazards and disaster reconnaissance. Frontiers in Built Environment, 6, p.573068. Galanis, Marios, et al. "DamageMap: A post-wildfire damaged buildings classifier." International Journal of Disaster Risk Reduction 65 (2021): 102540. Cao, Q.D. and Choe, Y., 2020. Building damage annotation on post-hurricane satellite imagery based on convolutional neural networks. Natural Hazards, 103(3), pp.3357–3376. Moradi, M., & Shah-Hosseini, R. (2020). Earthquake damage assessment based on deep learning method using VHR images. Environmental Sciences Proceedings, 5(1), 16. Rahnemoonfar, Maryam, et al. "Floodnet: A high resolution aerial imagery dataset for post flood scene understanding." IEEE Access 9 (2021): 89644–89654. Pi, Y., Nath, N.D. and Behzadan, A.H., 2021. Detection and semantic segmentation of disaster damage in UAV footage. Journal of Computing in Civil Engineering, 35(2), p.04020063. Deng, Liwei, and Yue Wang. "Post-disaster building damage assessment based on improved U-Net." Scientific Reports 12.1 (2022): 15862. Irwansyah, E., Young, H. and Gunawan, A.A., 2023. Multi disaster building damage assessment with deep learning using satellite imagery data. International Journal of Intelligent Systems and Applications in Engineering, 11(1), pp.122–131. Xia, Haobin, et al. "A Deep Learning Application for Building Damage Assessment Using Ultra-High-Resolution Remote Sensing Imagery in Turkey Earthquake." International Journal of Disaster Risk Science 14.6 (2023): 947–962. Fischer, E. et al. (2022) The 2021 Marshall Fire, Boulder County, Colorado. Available at: https://geerassociation.org/components/com_geer_reports/geerfiles/GEER Marshall Fire Report_v1_May 2022.pdf (Accessed: 01 August 2024). Gabbert, B. (2022) Report released for the Marshall Fire which destroyed 1,056 structures southeast of Boulder, Colorado, Wildfire Today. Available at: https://wildfiretoday.com/2022/10/29/report-released-for-the-marshall-fire-which-destroyed-1056-structures-southeast-of-boulder-colorado/ (Accessed: 01 August 2024). Computer vision tools for developers and enterprises Roboflow. Available at: https://roboflow.com/ (Accessed: 01 August 2024). Insurance Institute for Business and Home Safety (2018) Hurricane harvey wind damage assessment report summary. Available at: https://ibhs.org/wp-content/uploads/member_docs/Hurricane-Harvey-Wind-Damage-Assessment-Report-Summary_IBHS.pdf (Accessed: 01 August 2024). Zheng, Z., Zhong, Y., Wang, J., Ma, A. and Zhang, L., 2021. Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man-made disasters. Remote Sensing of Environment, 265, p.112636. Wang, Y., Chew, A.W.Z. and Zhang, L., 2022. Building damage detection from satellite images after natural disasters on extremely imbalanced datasets. Automation in construction, 140, p.104328. Asad, M.H., Asim, M.M., Awan, M.N.M. and Yousaf, M.H., 2023, March. Natural disaster damage assessment using semantic segmentation of uav imagery. In 2023 International Conference on Robotics and Automation in Industry (ICRAI) (pp. 1–7). IEEE. Wang, Y., Jing, X., Chen, W., Li, H., Xu, Y. and Zhang, Q., 2023. Geometry-informed deep learning-based structural component segmentation of post-earthquake buildings. Mechanical Systems and Signal Processing, 188, p.110028. Ahmadi, S.A., Mohammadzadeh, A., Yokoya, N. and Ghorbanian, A., 2024. BD-SKUNet: Selective-kernel UNets for building damage assessment in high-resolution satellite images. Remote Sensing, 16(1), p.182. Ronneberger, O., Fischer, P. and Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5–9, 2015, proceedings, part III 18 (pp. 234–241). Springer International Publishing. Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B. and Glocker, B., 2018. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999. Huang, H., Lin, L., Tong, R., Hu, H., Zhang, Q., Iwamoto, Y., Han, X., Chen, Y.W. and Wu, J., 2020, May. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP 2020–2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) (pp. 1055–1059). IEEE. Elharrouss, O., Akbari, Y., Almaadeed, N. and Al-Maadeed, S., 2022. Backbones-review: Feature extraction networks for deep learning and deep reinforcement learning approaches. arXiv preprint arXiv:2206.08016. He, K., Zhang, X., Ren, S. and Sun, J., 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770–778). Ioffe, S. and Szegedy, C., 2015, June. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning (pp. 448–456). pmlr. Huang, G., Liu, Z., Van Der Maaten, L. and Weinberger, K.Q., 2017. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700–4708). Tan, M. and Le, Q., 2019, May. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning (pp. 6105–6114). PMLR. Christoffersen, P. and Jacobs, K., 2004. The importance of the loss function in option valuation. Journal of Financial Economics, 72(2), pp.291–318. Lin, T.Y., Goyal, P., Girshick, R., He, K. and Dollár, P., 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (pp. 2980–2988). Yacouby, R. and Axman, D., 2020, November. Probabilistic extension of precision, recall, and f1 score for more thorough evaluation of classification models. In Proceedings of the first workshop on evaluation and comparison of NLP systems (pp. 79–91). Kingma, D.P. and Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Hussain, J. (2019). Deep Learning Black Box Problem (Dissertation). Retrieved from https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-393479 Algiriyage, N., Prasanna, R., Stock, K., Doyle, E.E. and Johnston, D., 2022. Multi-source multimodal data and deep learning for disaster response: a systematic review. SN Computer Science, 3, pp.1–29. Kang, D.K., Fischer, E., Olsen, M.J., Adams, J.A. and O’Neil-Dunne, J., 2024. Optimising disaster response: opportunities and challenges with Uncrewed Aircraft System (UAS) technology in response to the 2020 Labour Day wildfires in Oregon, USA. International Journal of Wildland Fire, 33(8). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 28 Aug, 2025 Reviews received at journal 06 Aug, 2025 Reviewers agreed at journal 01 Aug, 2025 Reviews received at journal 31 Jul, 2025 Reviewers agreed at journal 29 Jul, 2025 Reviewers agreed at journal 29 Jul, 2025 Reviewers agreed at journal 27 Jul, 2025 Reviewers agreed at journal 22 Jun, 2025 Reviews received at journal 16 Jun, 2025 Reviewers agreed at journal 11 Jun, 2025 Reviewers invited by journal 11 Jun, 2025 Editor assigned by journal 11 Jun, 2025 Editor invited by journal 09 Jun, 2025 Submission checks completed at journal 06 Jun, 2025 First submitted to journal 06 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6567339","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":470033974,"identity":"2efd2f7d-a9f4-43fa-9211-e66688f8ea13","order_by":0,"name":"Dae Kun Kang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIiWNgGAWjYBACCRDBw8AghybORliLMUKAjUgtiQ1Ea5FsP3v4xZsam/R+6cOHX3zcUVfHP7/5AMOHssM4tUjz5KVZzjmWljuzLy3NcuaZwxISx9gSGGecw61FjiHHzJiH7XDuhjM8Zsa8bQckGI7xGDDztuHRwv8GqOXf/3SDM/zfgFrqJOSP8X9g/otHi7REjvFjoOEJBmd4mIEMZgmDYzwMzIx4tEjOeGPGOLcv2XBmD5sZ48y2w5Ibj6UZHOw5l45Ti8T5HOMPb77ZyfMDLfnwsa2OX+7w4YcPfpRZ49QCBGwS6AyGA/jUAwHzB3TGKBgFo2AUjAIUAACYnVP+RLfiNQAAAABJRU5ErkJggg==","orcid":"","institution":"Oregon State University","correspondingAuthor":true,"prefix":"","firstName":"Dae","middleName":"Kun","lastName":"Kang","suffix":""},{"id":470033975,"identity":"94687e93-f64c-4ae0-bc7a-013174d66a6f","order_by":1,"name":"Michael J. Olsen","email":"","orcid":"","institution":"Oregon State University","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"J.","lastName":"Olsen","suffix":""},{"id":470033976,"identity":"c471e7cf-f162-4f97-9bcc-cecc840c1983","order_by":2,"name":"Erica Fischer","email":"","orcid":"","institution":"Oregon State University","correspondingAuthor":false,"prefix":"","firstName":"Erica","middleName":"","lastName":"Fischer","suffix":""},{"id":470033979,"identity":"bca1c306-7a92-4ccc-85c5-f0c324a14e0e","order_by":3,"name":"Jaehoon Jung","email":"","orcid":"","institution":"Gyeongsang National University","correspondingAuthor":false,"prefix":"","firstName":"Jaehoon","middleName":"","lastName":"Jung","suffix":""},{"id":470033980,"identity":"477725b8-5b3e-4226-bb88-e8775d7f225d","order_by":4,"name":"Julie A. Adams","email":"","orcid":"","institution":"Oregon State University","correspondingAuthor":false,"prefix":"","firstName":"Julie","middleName":"A.","lastName":"Adams","suffix":""}],"badges":[],"createdAt":"2025-04-30 18:38:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6567339/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6567339/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84570292,"identity":"5c6ff5aa-2774-4b35-b303-f7838070b9a9","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":556950,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.1 Houses damaged by Wildfire and Hurricane\u003c/p\u003e","description":"","filename":"4.1.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/55d78f3898cd06c71561e59f.png"},{"id":84570291,"identity":"562d09eb-c72c-4bc2-82fb-feaa370c0987","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":130712,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.2 Overview of the interoperability experiment process using Wildfire(W) and Hurricane(H) datasets\u003c/p\u003e","description":"","filename":"4.2.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/97741cde33e02739fb5ecfed.png"},{"id":84571209,"identity":"882cce68-bc84-4d25-b15d-f7a9f770044b","added_by":"auto","created_at":"2025-06-13 15:29:56","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":279177,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.3 Images and masks pairs of wildfire damaged buildings\u003c/p\u003e","description":"","filename":"4.3.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/280c801dc8772a15ff96b3e4.png"},{"id":84570296,"identity":"f6288c99-0ae4-4376-acf1-d9a050a7ee1d","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":297820,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.4 Images and masks pairs of hurricane damaged buildings\u003c/p\u003e","description":"","filename":"4.4.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/cbe0c6b6f66cc1d82edb43b7.png"},{"id":84570293,"identity":"65d16393-0ae2-496e-aa75-a063bfe68825","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":64779,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.5 Preprocessing of datasets (3,789 wildfire images and 4,494 hurricane images)\u003c/p\u003e","description":"","filename":"4.5.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/bfe2e586bee77f18c42f727f.png"},{"id":84570477,"identity":"69fc97df-1088-41b1-b2bc-c2054dc342ac","added_by":"auto","created_at":"2025-06-13 15:21:56","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":28547,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.6 Overview of model decision for different types of disaster\u003c/p\u003e","description":"","filename":"4.6.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/f5244df32809e2a187b5f42d.png"},{"id":84570479,"identity":"397c5f16-2ec7-4520-bc9c-75a126bc74e0","added_by":"auto","created_at":"2025-06-13 15:21:56","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":40403,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.7 Process to assess the cross-applicability of different types of disaster\u003c/p\u003e","description":"","filename":"4.7.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/9147a8f7fe75e19190a83515.png"},{"id":84570299,"identity":"192860d2-5715-44d1-b710-a96fc62b1264","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":29157,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.8 Process to assess the cross-applicability of different types of disaster models with transfer learning and various training data size\u003c/p\u003e","description":"","filename":"4.8.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/2bb1ea26896ca41386231dd7.png"},{"id":84570481,"identity":"32a15a4d-a136-4e44-a587-27a21e89810b","added_by":"auto","created_at":"2025-06-13 15:21:56","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":403489,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.9 Example results for the wildfire model, hurricane model, and combined model applied to the wildfire test dataset. (a) Wildfire test images; (b) Wildfire test mask; (c) W model; (d) H model; (e) W+H model\u003c/p\u003e","description":"","filename":"4.9.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/8d6e3c0d02c3f0ca12ae51fc.png"},{"id":84570307,"identity":"4079c43b-4391-427c-a6de-68907bd4dfc4","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":439694,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.10 Example results for the hurricane model, wildfire model, and combined model on the wildfire test dataset. (a) hurricane test images; (b) hurricane test mask; (c) H model; (d) W model; (e) W+H model\u003c/p\u003e","description":"","filename":"4.10.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/1f9874cc537c175fe31cb72f.png"},{"id":84570485,"identity":"f5707f6c-7f4f-42f8-a5c8-99c1b4803f7c","added_by":"auto","created_at":"2025-06-13 15:21:56","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":418172,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.11 Differences in the characteristics of debris of buildings affected by wildfire and hurricane\u003c/p\u003e","description":"","filename":"4.11.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/c07d2fba13a2ed1b7504ae58.png"},{"id":84570316,"identity":"ff457c80-cc63-404a-9709-3d18573d15ca","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":367280,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.12 Example results of the W model applied to the wildfire and hurricane test datasets highlighting the differences in debris distribution.\u003c/p\u003e","description":"","filename":"4.12.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/be87b2fb530b9ca0b1603c5b.png"},{"id":84571211,"identity":"a5cb62d3-7509-4425-8ec3-447b31ef2cf2","added_by":"auto","created_at":"2025-06-13 15:29:56","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":87778,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.13 The F1 score of the optimal H model applied to the W test dataset depending on whether transfer learning was applied and the amount of W training data.\u003c/p\u003e","description":"","filename":"4.13.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/c193ad220637add0a416266a.png"},{"id":84570308,"identity":"7e26e203-ab66-40db-b4de-c5c4b18df15f","added_by":"auto","created_at":"2025-06-13 15:13:56","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":88290,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4.14 The F1 score of the optimal W model applied to the H test dataset depending on whether transfer learning was applied and the amount of H training data used.\u003c/p\u003e","description":"","filename":"4.14.png","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/1432d7cfb7d4db5402361c0a.png"},{"id":84572164,"identity":"63e689e6-8a91-4ab3-bda3-894b3bd344a2","added_by":"auto","created_at":"2025-06-13 15:45:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4699651,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6567339/v1/23e51aec-2c1c-4e0e-a219-7342bb93d874.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep learning cross-applicability of Uncrewed Aircraft System (UAS) imagery from different disaster types for building damage assessment","fulltext":[{"header":"Introduction and background","content":"\u003cp\u003eNatural disasters, including wildfires, hurricanes, tornadoes, earthquakes, and floods, are prevalent globally, and increasing in frequency [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. For example, in the United States, 28 weather and climate disasters with losses exceeding USD 1\u0026nbsp;billion dollars in 2023 breaking the previous record of 22 set in 2020 [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. These disasters incurred a minimum cost of USD 92.9\u0026nbsp;billion [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Building damage accounts for a large portion of these costs. Identifying damage to buildings provides one of the costs associated with the disaster and also provides the basis for identifying affected population and determining an appropriate response [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. However, the collation of disaster-related information in the field is time-consuming and labor-intensive and often subjective [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Remote sensing technology from satellites and aircraft has provided timely, visual data about disaster areas, to assist with damage assessment. While more efficient and less subjective compared with manual methods, labeling and analyses of these data are time consuming. By analyzing the visual elements embedded in these data, computer vision technologies such as deep learning have come to play a pivotal role in addressing these shortcomings in post-disaster building damage assessment, especially when combined with rapidly advancing remote sensing technologies [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eLarge databases of labeled images are necessary to support these efforts. For example, the xBD dataset is a building damage assessment dataset containing pre- and post-event satellite images of disaster-related events (earthquake, tsunami, flood, wildfire, hurricane, and volcanic eruption) [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. To date, apart from the xBD dataset, there is no other integrated dataset that encompasses a wide range of disasters for post disaster building damage assessment. Although, there are some event specific datasets available from post reconnaissance activities [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], most of these have not been labeled to support deep learning analyses.\u003c/p\u003e \u003cp\u003eCurrent research has focused on post disaster building assessment with deep learning and remote sensing data concentrates on singular disaster types rather than looking at disasters holistically. For example, in exploring wildfire damage, Galanis et al. (2021) used a subset of the xBD dataset to apply ResNet (residual network), a convolutional neural network (CNN), to satellite images to classify buildings as damaged and undamaged [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. As another example, Cao and Choe (2020) applied a fine-tuned Visual Graphics Group (VGG)-16 to satellite images containing hurricane-damaged buildings and classified the buildings into damaged and undamaged categories [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. These studies showed high accuracy in classifying buildings damaged by each disaster, 93% in wildfire and 97% in hurricane. However, they did not provide pixel-level classification, which could provide more detailed information about the disaster area. Additionally, their performance was not verified on high-resolution images because they trained and tested based on satellite images. Moradi and Shah-Hosseini (2020) classified earthquake-damaged buildings into damaged and undamaged buildings using Very High Resolution (VHR) satellite images of Haiti and UNet [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. They provided pixel-level classification but used images before and after the earthquake simultaneously to improve accuracy. This has limitations in that high resolution images before the disaster are not always available and the image processing time is longer. Rahnemoonfar et al. (2021) classified flooded buildings using deep learning through a pixel-level analysis using several semantic segmentation modules (PSPNet, Enet, and DeepLabv3 plus) [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Pi et al. (2021) also provided pixel-wise analysis of flooded buildings in flood scenes but utilized video footage rather than static images [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlthough most previous studies address a single disaster, they provide the groundwork to advance their models and methodologies to be applicable to a variety of disasters as a future research direction. Recent studies have tried to integrate data across multiple disasters during model training to improve generalization capability. Deng et al. (2022) trained a deep learning model using data from multiple disasters in the xBD satellite dataset simultaneously [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. They divided the building damage classification into two stages building segmentation using pre-disaster images with UNet, and classification using pre- and post-disaster images. This method showed high accuracy of 87% in building localization and 75% in classification. Irwansyah et al. (2023) used the xBD dataset and performed building damage assessment through two steps: building localization with UNet and PSPNet and classification with ResNet50 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. However, since they compared the accuracy using only one backbone for UNet and PSPNet, there is potential for improvement by applying several backbones. Xia et al. (2023) proposed a BDANet model that integrates UNet and Cross-directional attention (CDA) module structure [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Although they did not apply their model to several disasters, they used the xBD dataset to train the model with images from multiple disasters. They achieved an average accuracy of 80% when applying the model to post-earthquake images from the Islahiye region of Turkey.\u003c/p\u003e \u003cp\u003eSeveral limitations arise in these prior works. First, many require both pre- and post-disaster images, which may not be available. Second, the classification task is performed in two stages, requiring substantial computation time. While semantic segmentation allows the original image to be used without preprocessing, the two-step approach fails to capitalize on this benefit. For example, extracting buildings as a preprocessing step undermines the efficiency that semantic segmentation offers. Third, models trained with xBD may have potential difficulties in scaling to high-resolution imagery given that xBD contains lower resolution satellite imagery that cannot capture details on damaged structures or debris piles that are important to the classification process.\u003c/p\u003e \u003cp\u003eAnother key challenge in the field of disaster damage assessment using deep learning is developing models that can be generalized and applied to multiple disaster types. However, there has been relatively little research on the mutual applicability of different disaster data types, which is fundamental for solving this problem. Investigating the mutual applicability of several types of disaster data can provide a basis for alleviating data scarcity, a fundamental issue in post-disaster building damage assessment using deep learning. By understanding the interrelationships between different disaster datasets, more effective data utilization can be achieved.\u003c/p\u003e \u003cp\u003eIn previous studies such as Irwansyah et al. (2023) [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], a vast amount of data from multiple disasters were simultaneously used for training; hence, it is difficult to identify the contribution of each disaster type. On the one hand, the use of additional images from other disaster types may substantially increase computational time with limited benefit to accuracy. Also, it is possible that the data from one disaster type may be detrimental to the classification of building damage from another disaster type given the different mechanisms. When categorizing data solely based on the damage extent the different damage mechanisms from each disaster type are neglected in \u0026lsquo;the training stage\u0026rsquo;. For instance, in a wildfire or tornado, most of the destroyed buildings will no longer be present at the site itself, with just the foundation and some debris remaining (Fig.\u0026nbsp;4.1a). In contrast, in a hurricane, destroyed buildings may have substantial roof damage but remain mostly intact or consist of a large debris pile (Fig.\u0026nbsp;4.1b).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(a) Wildfire affected houses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(b) Hurricane affected houses\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFigure 4.1 Houses damaged by Wildfire and Hurricane\u003c/p\u003e \u003cp\u003eThese observations raise a critical question central to this research: Are the differences in damage mechanisms and visual patterns across disaster types substantial enough to hinder the effective cross-application of data? In other words, despite the potential benefits of alleviating data scarcity, can data from one disaster type truly support damage assessment in another, or do these inherent disparities ultimately limit mutual applicability?\u003c/p\u003e \u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Research Objectives\u003c/h2\u003e \u003cp\u003eIn this study, high-resolution images acquired from UAS and semantic segmentation models were used to evaluate building damage at the pixel level for efficient and accurate damage assessment in complex disaster environments. In addition, through two datasets, wildfire and hurricane, we investigated the mutual applicability between disasters in a deep learning environment from two perspectives: the dataset and the model. The objectives of this study are (1) investigate the interrelationship between wildfire and hurricane data to explore the influence of different disaster data on analyzing buildings damaged by each disaster; (2) investigate the impact of the type and combination of backbones and models on accuracy; and (3) evaluate the inter-applicability of deep learning models developed from wildfire and hurricane data through transfer learning to extend a model developed from one disaster type to another. In this process, the effectiveness of the model relative to the amount of additional training data used is analyzed. Ultimately, this study contributes by providing a framework to demonstrate how to expand the data available for post-disaster building damage assessment by integrating different disaster datasets and can be extended to other disaster types.\u003c/p\u003e \u003c/div\u003e"},{"header":"Methods","content":"\u003cp\u003eThis study consists of four major parts (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e). First, in the data preparation stage, the wildfire and hurricane datasets to be used in this study were prepared with pixel-level annotations according to the degree of damage to buildings (standing and destroyed) and the characteristics of building damage from each disaster. Then, the backbones and deep learning models, loss function, accuracy assessment method, and the model training environment used to analyze these datasets were prepared. Second, several combinations of backbones and models were tested to find a deep learning model that is well suited for both wildfire and hurricane datasets. Third, the selected deep learning model was trained using the wildfire dataset, hurricane dataset, and the wildfire and hurricane integrated dataset. This model was then applied to each disaster dataset to analyze the impact of wildfire and hurricane data on each other in the deep learning environment. Fourth, the model developed from one disaster was applied to another disaster using transfer learning. The amount of training data used was varied and the associated sensitivity to the amount of training data was analyzed.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Datasets\u003c/h2\u003e \u003cp\u003eTo illustrate the framework, we utilize high-resolution Uncrewed Aircraft System (UAS) datasets obtained from two disaster types: wildfire and hurricane. These datasets consist of UAS images and masks containing pixel-level annotations of buildings for semantic segmentation annotation. Buildings are classified into standing and destroyed.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Wildfire dataset\u003c/h2\u003e \u003cp\u003eWildfire images were obtained from the Marshall fire, which burned over 6,080 acres and destroyed 1,056 structures in Boulder County, Colorado on December 30, 2021 [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Two UAS systems were utilized. First, a DJI Matrice 210 V2 RTK with X5S camera, a medium-sized UAS, captured high resolution imagery for a small town close to Mulberry St. Second, a Sensefly eBee X, a larger UAS with a flight endurance of up to one hour, was used to map large areas across the study area.\u003c/p\u003e \u003cp\u003eThe Roboflow platform was employed for the annotation of dataset for semantic segmentation [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. All buildings were manually digitized as a polygon along the building boundary. Three expert annotators were introduced, and the annotations were checked and revised by the authors before final approval. FEMA guidelines classify building damage into four levels: Affected, Minor, Major, and Destroyed. This generalized classification encompasses all disasters other than floods. However, most buildings in areas affected by wildfire generally are either completely burned down or show minimal damage. As a result, we classified the building damage into two categories. Standing, which referred to buildings with no observed damage or buildings requiring minimal repairs to make them habitable and destroyed, which referred to buildings requiring extensive repairs or buildings where the structural framing of the roof or walls has been damaged, leaving the interior exposed. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e shows the image and mask pairs of wildfire damaged buildings. Green represents standing buildings and red represents destroyed buildings.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Hurricane dataset\u003c/h2\u003e \u003cp\u003eRescueNet [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], contains detailed imagery captured following Hurricane Michael by the Center for Robot-Assisted Search and Rescue using small UAS (DJI Mavic Pro quadcopters) at Mexico Beach and other directly impacted areas. A total of 80 flights were performed between October 11\u0026ndash;14, 2018. Hurricane Michael made landfall on October 10, 2018, near Panama City, Florida, with sustained winds of approximately 160 mph. This event caused upwards of 25\u0026nbsp;billion dollars in damages with over 54,000 structures damaged or destroyed [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe original RescueNet dataset classified buildings with hurricane damage into four levels: Building-No-Damage, Building-Medium-Damage, Building-Major-Damage, and Building-Total-Destruction. However, to directly compare with the wildfire datasets in this study the categories of Building-No-Damage and Building-Medium-Damage were integrated into Standing and Building-Major-Damage and Building-Total-Destruction were integrated into Destroyed (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4.4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Preparation\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4.5\u003c/span\u003e provides an overview of the dataset preparation process for this study. In total, we prepared 3,789 wildfire images and 4,494 hurricane images with pixel-level annotations for buildings. Images were patched into 256x256 segments. These patches were divided 70:15:15 for deep learning model training, validation, and testing. Note that to maintain balance and avoid overfitting to a specific class, the division maintained a consistent ratio of standing buildings and destroyed buildings in each subset both in context of the number of polygons and pixels representing structures. Next, data augmentation was applied to these data through nondestructive transformations (e.g., rotate 90, 180, and 270 degrees, horizontal, vertical, and diagonal reflection) to maintain the information of the original image and mask. In this study, the wildfire dataset is represented as W dataset, and the wildfire test dataset used for model performance validation is represented as W test dataset. The hurricane dataset is also represented similarly: H dataset and H test dataset.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model training\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1 Semantic segmentation network architectures\u003c/h2\u003e \u003cp\u003eU-Net-based architectures have been used and proven in the field of post-disaster building damage assessment [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The U-Net, Attention U-Net, and U-Net 3 plus were explored for the base model (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e4.1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eU-Net is a convolutional neural network designed for image segmentation, developed by Ronneberger et al. in 2015 [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Its U-shaped architecture includes a contracting path (encoder) and an expansive path (decoder). The U-Net architecture is distinguished by its ability to extract image features and enable precise localization simultaneously. This is achieved through the integration of both low-dimensional and high-dimensional information. A key mechanism in U-Net is the use of skip connections, where features from each layer in the encoder are directly merged with the corresponding layer in the decoder. This concatenation method enhances the model\u0026rsquo;s capacity to retain spatial information and improve segmentation accuracy. U-Net's skip connections and precise localization capabilities allow it to accurately identify both the shapes of the entire building and the details of the damage.\u003c/p\u003e \u003cp\u003eAttention U-Net enhances the standard U-Net architecture by incorporating attention mechanisms the network to focus on relevant parts of the input image, thereby improving segmentation performance, particularly in cases involving small or complex structures [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The architecture introduces attention gates (AGs) into the standard U-Net design. AGs are employed during the skip connections procedure to focus on relevant features from the encoder before merging them with the decoder's features. These take features from both the encoder and the decoder, using them to generate an attention map that highlights important regions while suppressing irrelevant ones. The attended features are then concatenated with the upsampled features from the decoder. This improved focus is expected to help the network better capture contextual information and segment complex damaged buildings more effectively.\u003c/p\u003e \u003cp\u003eU-Net 3 plus builds upon the U-Net and U-Net 2 plus architectures by introducing full-scale skip connections, aiming to make full use of multi-scale features for more accurate segmentation [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. U-Net 3 plus redesigns the interconnections between the encoder and decoder, as well as the interconnections between decoder sub-networks. This enables the capture of fine-grained details and coarse-grained semantics across all scales. Each decoder layer incorporates feature maps from smaller- and same-scale encoder layers, and larger-scale decoder layers, merging them to enhance the segmentation performance. This approach is expected to be effective for buildings appearing at varying scales.\u003c/p\u003e \u003cp\u003eHowever, the accuracy of the models may vary depending on the combination of various backbones. A backbone refers to the established architecture or network employed for feature extraction, having been trained on numerous tasks previously and proven its efficacy [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. The following backbone networks were used in this study: Residual Neural Network (ResNet), Densely Connected Convolutional Network (DenseNet), and EfficientNet (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e4.1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eResNet is a model based on convolutional neural networks (CNNs), featuring the introduction of residual networks [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. ResNet has various versions depending on the number of parameter layers. ResNet includes skip connections or recurrent units between blocks of convolutional and pooling layers. Additionally, each block is followed by batch normalization [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. ResNet-101 with 44.7M parameters and ResNet-152 with 60.4M parameters are adopted in this study.\u003c/p\u003e \u003cp\u003eDenseNet is CNN architecture designed to enhance the flow of information and gradients throughout the network by creating direct connections between all layers [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. This facilitates the training of deeper and more complex networks. In DenseNet, each layer takes inputs from all previous layers, promoting feature reuse and improving gradient propagation, which leads to more efficient and effective training. DenseNet models tend to be more compact and require fewer parameters compared to traditional CNNs, making them ideal for deployment in environments with limited resources. DenseNet-169 with 14.3M parameters and DenseNet-201 with 20.2M parameters are adopted in this study.\u003c/p\u003e \u003cp\u003eEfficientNet is a modern family of architectures, showing outstanding performance in classification tasks while utilizing fewer parameters and Floating Point Operation (FLOP) than other networks [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. It employs compound scaling to efficiently and uniformly adjust the network's width, depth, and resolution. The various versions of EfficientNet, from B0 to B7, can be chosen based on resource availability and computational cost. EfficientNet-B4 with 19.5M parameters and EfficientNet-B7 with 66.7M parameters are adopted in this study.\u003c/p\u003e \u003cp\u003eIn this study, we experimentally developed models that is well-applied to wildfire and hurricane UAS datasets through the combination of three types of U-Net architectures (U-Net, Attention U-Net, and U-Net 3 plus) and six backbones (ResNet-101, ResNet-152, DenseNet-169, DenseNet-201, EfficientNet-B4, and EfficientNet-B7).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2 Loss function\u003c/h2\u003e \u003cp\u003eThe loss represents the total errors from each batch in the training or validation sets, indicating how well or poorly the trained model performs following each optimization step [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. In choosing the loss function, we focused on data imbalance both between the pixels of the standing building and the destroyed building and between these categories and the background pixels. We used the focal loss function (Categorical Focal Cross entropy) proposed by Lin et al. (2017) where the number of pixels are weighted inversely, thereby reducing the influence of the background, which occupies a large portion of each image [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of deep learning models and backbone networks used in the study\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eArchitectures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDetails\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eDeep Learning Models\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eU-Net [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- CNN for image segmentation with U-shaped architecture.\u003c/p\u003e \u003cp\u003e- Contracting (encoder) and expansive (decoder) paths.\u003c/p\u003e \u003cp\u003e- Skip connections for precise localization.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAttention U-Net [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- Enhanced U-Net with attention mechanisms.\u003c/p\u003e \u003cp\u003e- Focus on relevant image parts for better segmentation.\u003c/p\u003e \u003cp\u003e- Improved performance on complex or small structures.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eU-Net 3 Plus [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- Build on U-Net and U-Net 2 Plus.\u003c/p\u003e \u003cp\u003e- Full-scale skip connections.\u003c/p\u003e \u003cp\u003e- Capture fine-grained details and coarse semantics across scales.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eBackbone Networks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- CNN with residual networks.\u003c/p\u003e \u003cp\u003e- Skip connections for efficient feature extraction.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- CNN architecture with direct connections between all layers.\u003c/p\u003e \u003cp\u003e- Compact design ideal for resource-limited environments.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e- Modern architecture family with outstanding classification performance.\u003c/p\u003e \u003cp\u003e- Utilize compound scaling to adjust network width, depth, and resolution.\u003c/p\u003e \u003cp\u003e- Fewer parameters and FLOPs.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3 Accuracy assessment\u003c/h2\u003e \u003cp\u003eTo evaluate the performance of our model, the following four metrics (accuracy, precision, recall, and F1 score) were calculated at the pixel level for each class (standing and destroyed) using Equations 4.1\u0026ndash;4.4 [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. These factors consider the number of True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}=\\:\\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{P}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEq 4.1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}=\\:\\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEq 4.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}=\\:\\frac{\\text{T}\\text{P}+\\text{T}\\text{N}}{\\text{T}\\text{P}+\\text{F}\\text{N}+\\text{F}\\text{P}+\\text{T}\\text{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEq 4.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{F}1=\\frac{2\\text{*}\\text{T}\\text{P}}{2\\text{*}\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{F}\\text{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEq 4.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003ePrecision (Eq.\u0026nbsp;4.1) describes how many standing or destroyed building pixels are detected as standing or destroyed. Recall (Eq.\u0026nbsp;4.2) describes how many of the actual standing or destroyed building pixels have been detected. Accuracy (Eq.\u0026nbsp;4.3) describes how many pixels have been correctly detected, regardless of whether they are destroyed or standing. The F1 score (Eq.\u0026nbsp;4.4) provides a measure of the harmonic mean of precision and recall, which can reduce bias resulting from data imbalance. Hence, the F1 score was the primary factor considered to determine the model to reduce the impact of the imbalance in disaster datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.3.4 Setup for model training\u003c/h2\u003e \u003cp\u003eKeras with the TensorFlow was employed as a framework with the adaptive moment estimation (Adam) optimization algorithm [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. During the 100 epochs of training, validation data was sent across the network, and its loss was estimated and monitored. The network was trained in 16 batches until it reached convergence at an initial learning rate of 1e-4. To avoid overfitting, training was stopped when the validation loss did not increase for 5 epochs and saved the model with the smallest validation loss. Epochs, learning rate, and batch size were experimentally determined through repeated testing. Because the datasets used in this study were very large, limitations in computing power and time were important factors in determining these variables.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Base model and backbone evaluation\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e shows the overall process to combine the different backbones and deep learning models to determine the most suitable model for each disaster. A total of 18 models, combining 6 backbones and 3 deep learning models, were applied to each disaster dataset prepared through the dataset preparation process. Through these accuracy assessment metrics (accuracy, precision, recall, and F1 score) computed for each model, the overall performance for each disaster was evaluated. To select the optimal combination of models for each disaster, each was ranked by its F1 score, and the model with the lowest sum of ranks was selected as the base model. This process results in a total of three model combinations: 1) The optimal W model: the model that performed best on the W dataset, 2) the optimal H model: the model that performed best on the H dataset, and 3) the base model: the model that performed well overall on both disasters. The base model was used to analyze the cross-applicability of different types of disaster datasets. The optimal W and H models were subsequently used for cross-applicability of different types of disaster models with transfer learning.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Cross-applicability of different types of disaster datasets\u003c/h2\u003e \u003cp\u003eThis section describes the method of investigating the interrelation between different disaster datasets in a deep learning environment as a fundamental study for the integrated use of different disaster data. By analyzing results using different disaster datasets for training both separately and together, we can examine how each type of disaster data influences the analysis of the other. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4.7\u003c/span\u003e shows the overall process for investigating cross-applicability of different disaster types. Each dataset was trained with the base model determined in the previous section. The three trained models (wildfire model (W model), hurricane mode (H model), and combined model (W\u0026thinsp;+\u0026thinsp;H model)) were then applied to the W and H test datasets to evaluate their performance by changing the training dataset while using the same base model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Transfer learning\u003c/h2\u003e \u003cp\u003eWhile the previous section investigated the mutual applicability between disaster datasets, this section focuses on whether the use of transfer learning is a more efficient and effective method to apply a model developed from one disaster to another type of disaster. In this section, we cross-applied the optimal W and H models to each other to investigate the accuracy according to whether transfer learning is applied and the amount of training data used in order to apply a model developed from one disaster to another disaster (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4.8\u003c/span\u003e). To apply the optimal H model to the W test dataset, we re-trained the optimal H model with 1%, 3%, 5%, 10%, 20%, 30%, 50%, 75%, 100% of the entire W training dataset. We then analyzed the sensitivity of the model according to the data size. This same experiment was conducted both with and without transfer learning. To apply the optimal W model to the H test dataset, this same procedure was followed.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Results and Discussion","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Base model decision\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e shows the statistics (precision, recall, accuracy, and F1-score) for all 18 combinations of the backbones and deep learning models on H and W datasets. From the results in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e, we were able to find notable trends. The accuracy difference according to the number of backbone parameter layers was not significant. Although the difference in the number of parameters between the two versions was 6M for DenseNet at the minimum and 47.2M for EfficientNet at the maximum, differences in the F1 score were not significant overall, ranging from 0.0003 to 0.0262. Notably, the higher version with more parameters did not always show higher accuracy.\u003c/p\u003e \u003cp\u003eThe accuracy difference according to the backbone was larger than the accuracy difference according to the model. The difference in F1 score of the three models using the same backbone was from 0.0067 to 0.0201 on the H dataset, and from 0.0067 to 0.0290 on the W dataset. On the other hand, the differences in F1 scores of the six variations of the same model ranged from 0.0412 to 0.0650 on the H dataset and from 0.0390 to 0.0999 on the W dataset. This suggests that modifying the backbone of an existing model in post-disaster building damage assessment can potentially enhance accuracy.\u003c/p\u003e \u003cp\u003eThe model accuracy was average 0.0846 higher on the W dataset than on the H dataset. This issue may arise because the damage patterns in the H dataset are more complex and diverse than those in the W dataset. For example, it is relatively easy to identify the structural boundaries of buildings damaged by wildfires compared to buildings damaged by hurricanes, which results in differences in the accuracy of building annotations.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation metrics of all combinations of the backbones and deep learning models on hurricane and wildfire datasets (Important numbers are in bold)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBackbone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003eAccuarcy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c10\" namest=\"c9\"\u003e \u003cp\u003eF1-Socre\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHurricane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWildfire\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHurricane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eWildfire\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eHurricane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eWildfire\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eHurricane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eWildfire\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8204\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9104\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7685\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8304\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9664\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9869\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7934\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8671\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8215\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8955\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7759\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9669\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7980\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8921\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.7838\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7748\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8483\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8686\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.7614\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8782\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8617\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9610\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9858\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7789\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8698\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8160\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8165\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8846\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9688\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9893\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8162\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8971\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8269\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8998\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9129\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9698\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9899\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8197\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e\u003cb\u003e0.9062\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eAttention U-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8275\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7756\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8879\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8244\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8715\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7831\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8877\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9678\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9866\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8024\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8795\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8058\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8810\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7436\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8443\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9633\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8622\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.7944\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8764\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7381\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.7383\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9619\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9781\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8014\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8386\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9114\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8077\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9706\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9896\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8229\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.9013\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8443\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9183\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7855\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9700\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9896\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8984\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eU-Net 3 plus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.7333\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8363\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8893\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9596\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9889\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8961\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8430\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7742\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8876\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9691\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8062\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8940\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8041\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8808\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8493\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9641\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7827\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8648\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8185\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8807\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7216\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8643\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9635\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9861\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.7659\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8724\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8570\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9158\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7875\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9712\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9902\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.8201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.9058\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8420\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9108\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9004\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9716\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.9901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e0.8308\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.9056\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e shows the ranking of the combinations of backbones and deep learning models based on F1 score. As a result, U-Net 3 plus with EfficientNet B7 showed the highest performance with 0.8308 in the H dataset (the optimal H model), and U-Net with EfficientNet B7 showed the highest performance with 0.9062 in the W dataset (the optimal W model). Based on the sum of the ranks, U-Net 3 plus with EfficientNet B7 was selected as the base model because it performed well overall on both datasets.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRanking of the combinations of backbones and deep learning models based on F1 score (Important numbers are in bold)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBackbone\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHurricane\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eWildfire\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSum\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTotal Rank\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7934\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8671\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7980\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8686\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7789\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8698\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8162\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8971\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8197\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.9062\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eAttention U-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8024\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8229\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8984\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eU-Net 3 plus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 101\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet 152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8062\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8940\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7827\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDenseNet 201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7659\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8724\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9058\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNet B7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.8308\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9056\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Cross-applicability of different types of disaster datasets\u003c/h2\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1 Application to the wildfire dataset\u003c/h2\u003e \u003cp\u003eThree types of trained models from the base model were developed: the W model, the H model, and W\u0026thinsp;+\u0026thinsp;H model. Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4.4\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e4.9\u003c/span\u003e show the results of applying all three models to the W test dataset. The W model achieved high performance on the wildfire test dataset, confirmed by an F1 score of 0.9527 for the standing building class, 0.8507 for the destroyed building class, and 0.9017 on average. Also, comparing the ground truth in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e4.9\u003c/span\u003eb and the result of applying the W model in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e4.9\u003c/span\u003ec, we can visually confirm that this model detects both standing and destroyed buildings well. However, the H model showed very poor performance on the W test dataset (average F1\u0026thinsp;=\u0026thinsp;0.2903), particularly for destroyed buildings (F1\u0026thinsp;=\u0026thinsp;0.0153). As Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e4.9\u003c/span\u003ed shows, the H model hardly detects buildings damaged by wildfire. However, the W\u0026thinsp;+\u0026thinsp;H model, trained on a combined dataset of W and H data, achieved similarly high performance to the W model on the W test dataset (average F1\u0026thinsp;=\u0026thinsp;0.8936).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation metrics for the wildfire model (W), hurricane model (H), and combined model (W\u0026thinsp;+\u0026thinsp;H) on the wildfire test dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuarcy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eW model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9604\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9452\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9925\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9527\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8476\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8507\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9071\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8964\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9902\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9017\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eH model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7708\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4463\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9453\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.5653\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0354\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.0097\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9445\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.0153\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.2280\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9396\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.2903\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eH\u0026thinsp;+\u0026thinsp;W model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9478\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9451\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9464\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8559\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8263\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8408\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8857\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9886\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8936\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2 Application to the hurricane dataset\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e4.5\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4.10\u003c/span\u003e present the results of applying the H model, the W model, and the W\u0026thinsp;+\u0026thinsp;H model to the H test dataset. The overall trends were similar to when the models were applied to the W test dataset. The H model achieved high performance with an average F1 score of 0.8261 on the H test dataset. The W model showed low performance with an F1 score of 0.4302, which is similar to when the H model was cross applied to the W test dataset. Nevertheless, the H\u0026thinsp;+\u0026thinsp;W model achieved a similar performance to the H model with an F1 score of 0.8268 on the H test dataset.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation metrics for the hurricane model (H), wildfire model (W), and combined model (W\u0026thinsp;+\u0026thinsp;H) on the hurricane test dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuarcy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eH model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9278\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8697\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8978\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7662\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7428\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.7543\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8470\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8063\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9712\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8261\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eW model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7486\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9264\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.7217\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.1661\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1191\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.1388\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4314\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4339\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9136\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.4302\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eH\u0026thinsp;+\u0026thinsp;W model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStanding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9749\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9015\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDestroyed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7907\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7172\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9683\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.7522\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8463\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8092\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9716\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8268\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3 Comparison\u003c/h2\u003e \u003cp\u003eThe H model, the W model, and the W\u0026thinsp;+\u0026thinsp;H model were applied to the W and H test datasets, respectively, and the performances were statistically and visually analyzed. The results experimentally demonstrated that the damage patterns of buildings affected by hurricanes and wildfires are distinguished and cannot be ignored in a deep learning environment. Due to the complexity of deep neural networks, pinpointing the exact factors behind these results was challenging [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Therefore, in this study, we conducted a comparison analysis using the predicted result images to investigate the causes. This approach allowed us to identify two notable characteristics. The first is the difference in the characteristics of the debris. Debris of buildings damaged by wildfire is dominated by a damage pattern in which they are chemically transformed due to the influence of intense heat. However, in the case of hurricanes, physical transformation is dominant. As Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e4.11\u003c/span\u003e shows, the debris of buildings damaged by hurricanes and by wildfires differ in shape and size, which may have made it difficult for the H model trained on hurricane data to detect the debris of buildings damaged by wildfires.\u003c/p\u003e \u003cp\u003eThe second factor is the distribution of debris. Most of the debris from buildings damaged by wildfire remain within the building boundary. In contrast, hurricane-induced debris often extends beyond the building boundary due to strong winds. This difference in debris distribution patterns can negatively affect the accuracy when the W model is applied to H data, as it may misinterpret debris beyond the building boundary as a destroyed building. This can be visually confirmed through the difference between Fig.\u0026nbsp;4.12a and b.\u003c/p\u003e \u003cp\u003eFigure 4.12 Example results of the W model applied to the wildfire and hurricane test datasets highlighting the differences in debris distribution.\u003c/p\u003e \u003cp\u003eDespite these differences, the W\u0026thinsp;+\u0026thinsp;H model achieved high performance on both disaster datasets. Paradoxically, these results highlight the importance of investigating the interrelationships between datasets when integrating diverse disaster data. Combining these datasets without understanding their interrelationships can lead to an increase in data volume that does not improve accuracy and, consequently, prolongs model training time. Notably, training the model using the combined dataset took twice as long as training the model using a single dataset.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Transfer learning\u003c/h2\u003e \u003cp\u003eIn the previous section, we analyzed the impact of cross-applying W and H datasets from the perspective of the datasets, while in this section, we analyzed it from the perspective of deep learning models. We investigate how to cross-apply the optimal single disaster models to each other, so that a model developed from one disaster can be used for another.\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e4.3.1 Application to the wildfire dataset\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;4.13 shows the accuracy (F1 score) of the optimal H model applied to the W test dataset. However, unlike the previous section where the H model was directly applied to the W test dataset, the optimal H model was trained using the pre-trained weights on the H dataset using transfer learning. Portions of the W dataset ranging from 1\u0026ndash;100% were utilized in the transfer learning process. Figure\u0026nbsp;13a and b shows an experiment with and another without transfer learning applied to examine its benefits.\u003c/p\u003e \u003cp\u003eThe experiments revealed several notable results. First, when transfer learning was used, high accuracy could be achieved with nominal additions to the training dataset size. For example, when the model was trained using 133 wildfire images (5%), the F1 score was 0.7504 with transfer learning, which was 0.2669 higher than without transfer learning. Nevertheless, the difference in accuracy becomes smaller as the amount of training data used increases (e.g., difference of 0.1759 for 20%, 0.0348 for 50%, and 0.0031 for 100%). Second, applying transfer learning led to a significant improvement in accuracy with only 1% (27 images) of the training data added. Adding this small amount of training data increased the average F1 score from 0.2835 to 0.6198 (a difference of 0.3363). This improvement was particularly pronounced in the destroyed building class where the F1 score increased from 0.0610 to 0.4793 (a difference of 0.4183). Compared with the highest F1 score (approximately 0.9), 0.6198 might seem low; however, in the early stages of a disaster response, the timeliness of the information can be more important than achieving the highest levels of accuracy. For instance, immediately after a disaster, it may not be necessary to predict damaged buildings with 90% accuracy. If we can assess the damage status of buildings with 60% accuracy across the entire disaster site, this level of accuracy might be sufficient to deploy safety personnel. Understanding that the required accuracy may vary over time, it is a meaningful discovery that we can achieve adequate accuracy by adding a small amount of training data specific to the disaster event as more data are collected and available.\u003c/p\u003e\u003cp\u003eFigure 4.13 The F1 score of the optimal H model applied to the W test dataset depending on whether transfer learning was applied and the amount of W training data.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e4.3.2 Application to the hurricane dataset\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;4.14 shows the accuracy (F1 score) of the optimal W model applied to the H test dataset depending on whether transfer learning was applied and with varying the amount of training data. The overall trend of the results of this experiment was similar to when the same experiment with the W dataset. When transfer learning was used, higher accuracy was achieved with the same amount of added training data, but the difference became smaller as the size of the training dataset increased. Similarly, a large increase in accuracy was achieved with a very small addition of training data. When trained using 1% (45 images) of the entire H dataset, the average F1 score improved by 0.1030, from 0.4146 to 0.5176, which is the largest improvement observed. However, this increase is much smaller compared to the improvement in the wildfire case, which was 0.4183. A notable difference is that the slope of the increase in accuracy relative to the amount of training data is more linear and gradual compared with the wildfire dataset. We hypothesize that the reason for these differences result from the fact that the characteristics of hurricane building damage are more diverse than those of wildfire.\u003c/p\u003e\u003cp\u003eFigure 4.14 The F1 score of the optimal W model applied to the H test dataset depending on whether transfer learning was applied and the amount of H training data used.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003e4.3.3 Comparison\u003c/h2\u003e \u003cp\u003eIn the previous section, we conducted experiments to see whether models developed in one disaster can be applied to other disasters using transfer learning. From the results, we were able to find notable trends. First, deep learning models showed higher performance in standing buildings than in destroyed buildings, regardless of whether transfer learning was used. This is because standing buildings are more clearly defined than destroyed buildings. Standing buildings have similar and regular shapes, patterns, and colors, which is advantageous in feature extraction and classification. On the other hand, destroyed buildings show a wide range of variations and irregularities depending on the degree of damage caused by the disaster. As a result, it is difficult to extract common features compared to standing buildings. However, we found that the performance difference between the two categories was large with a small number of training data, but as the number of training data increased, the performance on destroyed buildings improved significantly, approaching the performance on standing buildings.\u003c/p\u003e \u003cp\u003eThe performance on standing buildings according to the increase in training data was different in hurricanes and wildfires. Comparing Fig.\u0026nbsp;4.13b and Fig.\u0026nbsp;4.14b, when using a small number of training data, the performance on standing buildings tends to be higher in hurricanes than in wildfires. This can be seen because the diversity of standing buildings in the dataset is higher in wildfires, which requires more types of data. However, as can be seen in Fig.\u0026nbsp;4.13a and Fig.\u0026nbsp;4.14a, it was found that the difference from the diversity of standing buildings, which may vary depending on the disaster or region, can be significantly improved through transfer learning.\u003c/p\u003e \u003cp\u003eFinally, the introduction of transfer learning led to a dramatic reduction in training time in both datasets. Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e shows the training time per epoch of the optimal models applied to the test dataset depending on whether transfer learning was applied and the amount of training data used. When transfer learning was applied, the optimal H model was shortened from 833s to 420s, and the optimal W model was shortened from 1067s to 573s, showing a time reduction of about 2 times in both cases. Also, in both cases, the amount of training data used and the training time per epoch are proportional. Combining this result with the accuracy from the previous section, we can experimentally see how much training data is required to train the model, the training time, and the corresponding accuracy. For example, if we apply the hurricane model to wildfire using transfer learning, it means that with about 1% (27 images) of the training data, we can achieve an accuracy of about 60% (average F1\u0026thinsp;=\u0026thinsp;0.6198) and about 660 seconds of training for 100 epochs. The results of these experiments clearly demonstrate the usefulness of transfer learning and its corresponding benefits in the field of disaster assessment.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTraining time per epoch of the optimal H and W model applied to the W and H test dataset depending on whether transfer learning was applied and the amount of W and H training data used (TL\u0026thinsp;=\u0026thinsp;with transfer learning; nTL\u0026thinsp;=\u0026thinsp;without transfer learning; H\u0026thinsp;=\u0026thinsp;the optimal H model applied to the W test dataset; W\u0026thinsp;=\u0026thinsp;the optimal W model applied to the H test dataset; Units\u0026thinsp;=\u0026thinsp;sec)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"11\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e10%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e20%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e30%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003e50%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003e75%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e126\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e210\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e315\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e420\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enTL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e167\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e833\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eW\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e115\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e172\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e287\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e430\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e573\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enTL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e107\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e213\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e320\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e533\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e1067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Challenges and future directions\u003c/h2\u003e \u003cp\u003eOne of the primary challenges of this study was the availability of insufficient high-resolution post-disaster datasets that captured building damage. We used datasets obtained from one wildfire (Marshall fire) and one hurricane (Hurricane Michael) event. Therefore, our experimental results may be overfitted to these specific datasets. The simplest solution would be to use datasets from a variety of events, but this is not easily achievable in practice. Efforts to apply deep learning to post-disaster building damage assessment are steadily increasing, especially for those containing high-resolution images [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. With the recent increase in the use of UAS in disaster response, we anticipate that more high-resolution disaster datasets will become available in the future [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. However, beyond the challenges in acquiring these data in the short window following a disaster, the annotation process required for deep learning is extremely labor-intensive, which can pose another major barrier. Hence, it is anticipated that annotation of disaster data will remain slow in the near future. In light of these challenges, we made the following efforts were made to minimize overfitting due to insufficient data. First, data augmentation was implemented to increase the diversity of the dataset. Additionally, early stopping was applied during the training process. Lastly, batch normalization was used to normalize the data within each batch based on the mean and variance, ensuring consistency across different distributions during the learning process.\u003c/p\u003e \u003cp\u003eAnother challenge is the absence of standards for systematic disaster data acquisition and processing. To develop models applicable to multiple types of disasters, the integrated use of diverse disaster data from multiple sources is essential. However, most post-disaster datasets currently acquired using UAS are collected and processed according to the criteria of the respective groups that gather them. These variations in flight pattern strategies, UAS shooting angles and altitudes, and desired resolutions can hinder data integration. While this diversity can eventually be beneficial in increasing data variety when sufficient data has become available, there is currently no benchmark dataset for post-disaster building damage assessment that includes high-resolution images. Establishing standards to facilitate systematic data collection, integration, processing, and labeling is necessary [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWith respect to the annotation process, there is also no common standard for classifying building damage across the different disaster datasets obtained with UAS or other remote sensing data source. Many studies follow FEMA's guidelines, which are based on ground observations and only categorize disasters into flood or others. Since many UAS collections consist of a nadir aerial view, the guidelines and categorization require adaptation. For example, when annotating a building damaged by a hurricane, where should the boundaries of the damaged building be set? Should building damage from all disasters be categorized into four categories or two? Establishing clear guidelines for annotating building damage will ensure consistency and improve the accuracy and reliability of models developed for post-disaster building damage assessment as multiple data sources are combined.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eAs foundational research to expand the data availability for post-disaster building damage assessment and to develop a model applicable to various disasters, this study explored the interoperability between different types of high-resolution disaster datasets, focusing on wildfire and hurricane datasets. By evaluating and combining several backbones and deep learning models, we successfully developed models to assess building damage caused by these disasters. In addition, we cross-applied these models to each disaster dataset to explore the possibility of integrating the two disaster datasets. The results experimentally demonstrated that the damage patterns of buildings affected by hurricanes and wildfires are distinct and cannot be ignored in a deep learning environment. Nevertheless, we demonstrated that the models developed from each disaster can be successfully cross-applied through transfer learning. A notable finding was that when cross applying a model developed from one disaster to another, a significant increase in accuracy can be achieved with a very small amount of additional training data when transfer learning is utilized. This accuracy increase was particularly evident on the wildfire dataset.\u003c/p\u003e \u003cp\u003eFuture studies will focus on expanding the diversity of disaster datasets by including data from several events and sources. Efforts will also be made to standardize the acquisition and processing of high-resolution post-disaster datasets. Furthermore, establishing standard guidelines for annotating building damage across different types of disasters is essential. Additionally, research will continue to explore methods to mitigate overfitting. By addressing these challenges, future work can improve the accuracy and reliability of models for post-disaster building damage assessment, making them applicable to a broader range of disaster scenarios.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank the NHERI RAPID Facility (CMMI- 2130997) and GEER team (CMMI-1266418) who helped in data collection and processing. The authors also thank Maryam Rahnemoonfar, Tashnim Chowdhury, and Robin Murphy who shared the hurricane dataset. Also,\u0026nbsp;the authors are grateful to the Federal Aviation Administration (FAA) and Alliance for System Safety of UAS through Research Excellence (ASSURE) for research\u0026nbsp;funding, which indirectly supported this research. Lastly,\u0026nbsp;the authors appreciate the support of Geoffrey Hollinger who provided helpful feedback on the manuscript.\u003c/p\u003e\n\u003cp\u003eAuthor information\u003c/p\u003e\n\u003cp\u003eAuthors and Affiliations\u003c/p\u003e\n\u003cp\u003eSchool of Civil and Construction Engineering, Oregon State University, Corvallis, OR, USA\u003c/p\u003e\n\u003cp\u003eDae Kun Kang, Michael J. Olsen, Erica Fischer\u003c/p\u003e\n\u003cp\u003eDepartment of Urban Engineering, Gyeongsang National University, Jinju, South Korea\u003c/p\u003e\n\u003cp\u003eJaehoon Jung\u003c/p\u003e\n\u003cp\u003eSchool of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR, USA\u003c/p\u003e\n\u003cp\u003eJulie A. Adams\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis work was supported by Oregon State University through graduate research assistantships. No other funding was received for conducting this study.\u003c/p\u003e\n\u003cp\u003eData Availability\u003c/p\u003e\n\u003cp\u003eThe datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003eEthics approval\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eConsent to participate\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests to disclose.\u003c/p\u003e\n\u003cp\u003eConsent to publish declaration\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eClinical trial\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eRitchie, H., Rosado, P. and Roser, M., 2023. Natural disasters. Our World in Data.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith, A.B. (2024) 2023: A historic year of U.S. billion-dollar weather and climate disasters, NOAA Climate.gov. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.climate.gov/news-features/blogs/beyond-data/2023-historic-year-us-billion-dollar-weather-and-climate-disasters\u003c/span\u003e\u003cspan address=\"https://www.climate.gov/news-features/blogs/beyond-data/2023-historic-year-us-billion-dollar-weather-and-climate-disasters\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Accessed: 05 August 2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDell'Acqua, F. and Gamba, P., 2012. Remote sensing and earthquake damage assessment: Experiences, limits, and perspectives. Proceedings of the IEEE, 100(10), pp.2876\u0026ndash;2890.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLallemant, D., Soden, R., Rubinyi, S., Loos, S., Barns, K. and Bhattacharjee, G., 2017. Post-disaster damage assessments as catalysts for recovery: A look at assessments conducted in the wake of the 2015 Gorkha, Nepal, earthquake. Earthquake Spectra, 33(1_suppl), pp.435\u0026ndash;451.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRahnemoonfar, M., Chowdhury, T. and Murphy, R., 2023. RescueNet: a high resolution UAV semantic segmentation dataset for natural disaster damage assessment. Scientific data, 10(1), p.913.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGupta, R., Goodman, B., Patel, N., Hosfelt, R., Sajeev, S., Heim, E., Doshi, J., Lucas, K., Choset, H. and Gaston, M., 2019. Creating xBD: A dataset for assessing building damage from satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (pp. 10\u0026ndash;17).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerman, J.W., Wartman, J., Olsen, M., Irish, J.L., Miles, S.B., Tanner, T., Gurley, K., Lowes, L., Bostrom, A., Dafni, J. and Grilliot, M., 2020. Natural hazards reconnaissance with the NHERI RAPID facility. Front Built Environ 6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWartman, J., Berman, J.W., Bostrom, A., Miles, S., Olsen, M., Gurley, K., Irish, J., Lowes, L., Tanner, T., Dafni, J. and Grilliot, M., 2020. Research needs, challenges, and strategic approaches for natural hazards and disaster reconnaissance. Frontiers in Built Environment, 6, p.573068.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGalanis, Marios, et al. \"DamageMap: A post-wildfire damaged buildings classifier.\" International Journal of Disaster Risk Reduction 65 (2021): 102540.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao, Q.D. and Choe, Y., 2020. Building damage annotation on post-hurricane satellite imagery based on convolutional neural networks. Natural Hazards, 103(3), pp.3357\u0026ndash;3376.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoradi, M., \u0026amp; Shah-Hosseini, R. (2020). Earthquake damage assessment based on deep learning method using VHR images. Environmental Sciences Proceedings, 5(1), 16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRahnemoonfar, Maryam, et al. \"Floodnet: A high resolution aerial imagery dataset for post flood scene understanding.\" IEEE Access 9 (2021): 89644\u0026ndash;89654.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePi, Y., Nath, N.D. and Behzadan, A.H., 2021. Detection and semantic segmentation of disaster damage in UAV footage. Journal of Computing in Civil Engineering, 35(2), p.04020063.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng, Liwei, and Yue Wang. \"Post-disaster building damage assessment based on improved U-Net.\" Scientific Reports 12.1 (2022): 15862.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIrwansyah, E., Young, H. and Gunawan, A.A., 2023. Multi disaster building damage assessment with deep learning using satellite imagery data. International Journal of Intelligent Systems and Applications in Engineering, 11(1), pp.122\u0026ndash;131.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXia, Haobin, et al. \"A Deep Learning Application for Building Damage Assessment Using Ultra-High-Resolution Remote Sensing Imagery in Turkey Earthquake.\" International Journal of Disaster Risk Science 14.6 (2023): 947\u0026ndash;962.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFischer, E. et al. (2022) The 2021 Marshall Fire, Boulder County, Colorado. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://geerassociation.org/components/com_geer_reports/geerfiles/GEER\u003c/span\u003e\u003cspan address=\"https://geerassociation.org/components/com_geer_reports/geerfiles/GEER\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e Marshall Fire Report_v1_May 2022.pdf (Accessed: 01 August 2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGabbert, B. (2022) Report released for the Marshall Fire which destroyed 1,056 structures southeast of Boulder, Colorado, Wildfire Today. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://wildfiretoday.com/2022/10/29/report-released-for-the-marshall-fire-which-destroyed-1056-structures-southeast-of-boulder-colorado/\u003c/span\u003e\u003cspan address=\"https://wildfiretoday.com/2022/10/29/report-released-for-the-marshall-fire-which-destroyed-1056-structures-southeast-of-boulder-colorado/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Accessed: 01 August 2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eComputer vision tools for developers and enterprises Roboflow. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://roboflow.com/\u003c/span\u003e\u003cspan address=\"https://roboflow.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Accessed: 01 August 2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInsurance Institute for Business and Home Safety (2018) Hurricane harvey wind damage assessment report summary. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ibhs.org/wp-content/uploads/member_docs/Hurricane-Harvey-Wind-Damage-Assessment-Report-Summary_IBHS.pdf\u003c/span\u003e\u003cspan address=\"https://ibhs.org/wp-content/uploads/member_docs/Hurricane-Harvey-Wind-Damage-Assessment-Report-Summary_IBHS.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Accessed: 01 August 2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng, Z., Zhong, Y., Wang, J., Ma, A. and Zhang, L., 2021. Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man-made disasters. Remote Sensing of Environment, 265, p.112636.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Y., Chew, A.W.Z. and Zhang, L., 2022. Building damage detection from satellite images after natural disasters on extremely imbalanced datasets. Automation in construction, 140, p.104328.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsad, M.H., Asim, M.M., Awan, M.N.M. and Yousaf, M.H., 2023, March. Natural disaster damage assessment using semantic segmentation of uav imagery. In 2023 International Conference on Robotics and Automation in Industry (ICRAI) (pp. 1\u0026ndash;7). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Y., Jing, X., Chen, W., Li, H., Xu, Y. and Zhang, Q., 2023. Geometry-informed deep learning-based structural component segmentation of post-earthquake buildings. Mechanical Systems and Signal Processing, 188, p.110028.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhmadi, S.A., Mohammadzadeh, A., Yokoya, N. and Ghorbanian, A., 2024. BD-SKUNet: Selective-kernel UNets for building damage assessment in high-resolution satellite images. Remote Sensing, 16(1), p.182.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRonneberger, O., Fischer, P. and Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015: 18th international conference, Munich, Germany, October 5\u0026ndash;9, 2015, proceedings, part III 18 (pp. 234\u0026ndash;241). Springer International Publishing.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B. and Glocker, B., 2018. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, H., Lin, L., Tong, R., Hu, H., Zhang, Q., Iwamoto, Y., Han, X., Chen, Y.W. and Wu, J., 2020, May. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP 2020\u0026ndash;2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) (pp. 1055\u0026ndash;1059). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElharrouss, O., Akbari, Y., Almaadeed, N. and Al-Maadeed, S., 2022. Backbones-review: Feature extraction networks for deep learning and deep reinforcement learning approaches. arXiv preprint arXiv:2206.08016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., Zhang, X., Ren, S. and Sun, J., 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770\u0026ndash;778).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoffe, S. and Szegedy, C., 2015, June. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning (pp. 448\u0026ndash;456). pmlr.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, G., Liu, Z., Van Der Maaten, L. and Weinberger, K.Q., 2017. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700\u0026ndash;4708).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan, M. and Le, Q., 2019, May. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning (pp. 6105\u0026ndash;6114). PMLR.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChristoffersen, P. and Jacobs, K., 2004. The importance of the loss function in option valuation. Journal of Financial Economics, 72(2), pp.291\u0026ndash;318.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin, T.Y., Goyal, P., Girshick, R., He, K. and Doll\u0026aacute;r, P., 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (pp. 2980\u0026ndash;2988).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYacouby, R. and Axman, D., 2020, November. Probabilistic extension of precision, recall, and f1 score for more thorough evaluation of classification models. In Proceedings of the first workshop on evaluation and comparison of NLP systems (pp. 79\u0026ndash;91).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma, D.P. and Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHussain, J. (2019). Deep Learning Black Box Problem (Dissertation). Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-393479\u003c/span\u003e\u003cspan address=\"https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-393479\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlgiriyage, N., Prasanna, R., Stock, K., Doyle, E.E. and Johnston, D., 2022. Multi-source multimodal data and deep learning for disaster response: a systematic review. SN Computer Science, 3, pp.1\u0026ndash;29.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang, D.K., Fischer, E., Olsen, M.J., Adams, J.A. and O\u0026rsquo;Neil-Dunne, J., 2024. Optimising disaster response: opportunities and challenges with Uncrewed Aircraft System (UAS) technology in response to the 2020 Labour Day wildfires in Oregon, USA. International Journal of Wildland Fire, 33(8).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-civil-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Civil Engineering](https://www.springer.com/journal/44290)","snPcode":"44290","submissionUrl":"https://submission.nature.com/new-submission/44290","title":"Discover Civil Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6567339/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6567339/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn recent years, the frequency and intensity of natural disasters have increased worldwide. These disasters cause significant economic losses, with building damage accounting for a substantial portion. Post disaster response plans involve inventories of building damage. However, these assessments can be highly subjective, require substantial time, and can expose the inspectors to unsafe environments. Automation of building damage assessment by applying deep learning combined with advanced remote sensing technology is currently an active research topic. These efforts are hindered by the limited amount of high-quality training datasets for each disaster type (e.g., hurricane, wildfire). Buildings damaged by different disasters may show distinct damage patterns due to differing damage mechanisms, posing challenges to data integration across disaster types and model development. To investigate these issues, this study explores the interrelationship between wildfire and hurricane data by developing models suited to wildfire and hurricane datasets both individually and jointly as well as combining various backbones and deep learning models. Our approach includes semantic segmentation for pixel-level damage assessment and analyzing model sensitivity with different amounts of training data. Ultimately, this study provides a solution to the limited data available to train building damage assessment deep learning models by providing a comparative analysis of the inter-applicability of wildfire and hurricane data. A notable finding is that when using a small portion of data through transfer learning, data and deep learning models from the other disaster types can be leveraged.\u003c/p\u003e","manuscriptTitle":"Deep learning cross-applicability of Uncrewed Aircraft System (UAS) imagery from different disaster types for building damage assessment","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-13 15:13:51","doi":"10.21203/rs.3.rs-6567339/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-08-28T10:54:13+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-06T20:54:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"261001929425323836382391743719418518598","date":"2025-08-01T10:11:18+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-31T17:18:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"317658647225021731099871412226574048385","date":"2025-07-29T14:39:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"221419263456358335579009187430112513847","date":"2025-07-29T13:09:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"64918391452459145040713022314700901852","date":"2025-07-27T14:18:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"315418802967870967412189428101441460933","date":"2025-06-23T02:26:04+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-16T19:21:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"113747018515434892244747752520392128799","date":"2025-06-11T22:46:47+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-11T10:13:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-11T09:59:41+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-06-09T16:31:26+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-06T16:00:53+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Civil Engineering","date":"2025-06-06T15:56:42+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-civil-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Civil Engineering](https://www.springer.com/journal/44290)","snPcode":"44290","submissionUrl":"https://submission.nature.com/new-submission/44290","title":"Discover Civil Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a6d8218e-f2db-46c6-8c2b-484beb103268","owner":[],"postedDate":"June 13th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-10-31T06:23:41+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-13 15:13:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6567339","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6567339","identity":"rs-6567339","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}