Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning

Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning | 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 Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning Jiqiang Xie, Bing Yu, Heng Lyu, Shengnan Fu, Chen Yang, Chi Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7476416/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Compound floods, driven by concurrent heavy rainfall and high tides, increasingly threaten coastal urban areas with severe infrastructure damage and socioeconomic disruption. Rapid identification of inundation extents and their dominant drivers is crucial for enabling timely emergency response and infrastructure protection. Current machine learning (ML) models excel at predicting flood inundation extents, but their 'black-box' nature restricts identifying dominant local drivers. While explainable AI (XAI) techniques are emerging to address this challenge, it is necessary to determine how to best pair an XAI method with a high-performing ML model to ensure both predictive accuracy and robust interpretability. This study systematically evaluates Explainable Machine Learning (EML) models to predict inundation areas and identify dominant drivers. Our EML models were created by pairing two representative XAI techniques, SHAP and LIME, with three distinct types of ML models: a linear model (Logistic Regression), a tree-based ensemble (Random Forest), and a neural network (Multilayer Perceptron). Norfolk, Virginia, USA was selected to train, test, and conduct driver analysis for the models. Results revealed that the RF achieved the best predictive performance (Accuracy = 0.81). Furthermore, of the two XAI techniques evaluated, SHAP-based driver attribution demonstrated greater consistency with real-world conditions because its ability to account for complex driver interactions, such as rainfall and tides, provides a more reliable identification of their influence. By leveraging XAI techniques, ML models can move beyond prediction to guiding informed decision-making and developing more effective flood management strategies. Urban flood Explainable machine learning Flood inundation areas Dominant driver Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Highlights A framework based on explainable machine learning models is proposed to predict flood maximum inundation extent and identify the dominant flood drivers. Systematic comparative analysis revealed that RF combined with SHAP performed the best for the quick and accurate identification of flood maximum inundation extent and their dominant drivers. SHAP outperforms LIME by accounting for both direct contributions and interactions of driving factors. 1 Introduction Compound floods, driven by multiple interacting factors, are often more destructive than single-driver events. These events frequently cause substantial property damage and loss of life (Wahl et al. 2015 ). The severe impacts of compound floods underscore the critical role of emergency response in protecting lives and property, particularly as flood patterns evolve (El baida et al. 2024 ). For example, in the United States, local Police and Fire Departments lead urban emergency flood responses, rapidly deploying resources to minimize harm (Yin et al., 2017). Similarly, the UK’s Environment Agency (EA) manages national flood emergency operations and deploying temporary defenses for numerous high-risk areas (Environment Agency, 2020). Such national efforts highlight that effective and timely emergency response is indispensable for urban flood resilience (Asif et al. 2025 ). Rapid identification of inundation extent and dominant drivers provides the crucial basis for guiding effective emergency response in urban areas prone to compound floods. Predicting inundation extent aids timely evacuations (Yin et al. 2024 ) and the deployment of temporary defenses (Qin et al., 2024a ). Equally crucial is the identification of flood dominant drivers. In compound floods, various drivers (e.g., heavy rainfall, high river levels, storm surges, tides) interact or combine, leading to diverse flood types (Fang et al., 2021 ). Because each flood type necessitates a tailored response strategy in practsice (Table 1 ), pinpointing the dominant driver—the primary factors determining the degree of flood inundation—serves to guide effective and appropriate interventions. For example, when inundation results from insufficient drainage capacity due to intense local rainfall, localized measures—such as opening stormwater inlets, excavating temporary channels, and constructing diversion pathways using temporary water barriers—can effectively mitigate surface flooding. Conversely, when inundation is driven by high sea levels from storm surges or excessive river flow, which overwhelm drainage systems and cause backflow, localized measures alone are insufficient. In such scenarios, external emergency interventions, such as deploying mobile pumps and water storage tanks to extract and drain accumulated surface water (Qin et al. 2024a ), as well as regulating downstream reservoirs or pumping stations to reduce river water levels, become essential. Thus, accurately identifying the dominant drivers of flood inundation is crucial for differentiating between locally manageable situations and those necessitating external support, thereby optimizing emergency resource allocation. Table 1 Types of floods and corresponding mitigation measures Causes of flooding Mitigation measures Literature Rainfall intensity exceeds the drainage capacity of the system Opening stormwater inlets, excavating temporary channels, constructing diversion pathways with temporary water barriers, mobile pumps and mobile water storage tanks (Qin et al. 2024b ) (Qin et al. 2024a ) (Shen et al. 2019 ) Upstream runoff exceeding the river's capacity Regulation of reservoirs or pumping stations to lower river water levels (Sun et al. 2023 ) (Zhao et al. 2022 ) Sea level rise and storm surges Regulation of tide gate, temporary water barriers (Shen et al. 2019 ) (Han and Tahvildari 2024 ) Flood inundation simulation typically relies on hydrodynamic models recognized for their accuracy and robust representation of underlying physical processes (Razavi et al. 2012 ). However, their significant computational demands often hinder practical application in real-time flood forecasting, despite various acceleration efforts (Rong et al. 2024 ; Sanders and Schubert 2019 ). In contrast, machine learning (ML) models, including deep learning structures, can rapidly learn complex input-output relationships from training data. This capability enables accurate flood inundation predictions (Tian et al. 2025 ). A spectrum of ML techniques has been investigated, from linear and tree-based algorithms to more sophisticated neural network architectures (Teng et al. 2017 ; Bentivoglio et al. 2022 ). For instance, Mangukiya et al.(2024) utilized the HEC-RAS framework to develop linear logistic regression, decision tree, and XGBoost models for simulating flood depth and inundated areas, reporting high accuracy and computational efficiency. Similarly, Kabir et al.(2020) employed a 1D Convolutional Neural Network for fluvial flood inundation modeling in Carlisle, UK, achieving prediction speeds 37 times faster than hydrodynamic model. These advancements highlight the potential of machine learning (ML) models as powerful tools for real-time flood inundation prediction. While machine learning models excel at the predictive tasks, their common 'black-box' nature poses a significant challenge when attempting to use them for the explicit identification of dominant drivers in specific inundated areas (Rudin 2019 ; Materia et al. 2024 ). Existing methods for driver analysis also have significant shortcomings. First, analytical techniques like feature contribution analyses (e.g., Partial Dependence Plots) are limited because they typically provide only global assessments of driver importance across an entire study area, failing to offer the location-specific insights crucial for targeted interventions (Zhou et al., 2024 ; Rohmer et al., 2018 ). Second, Physics-based methods, including hydrodynamic modeling combined with scenario analysis, are also commonly used to identify dominant flood drivers. Although these methods can effectively attribute flooding to different drivers, they are computationally intensive, limiting their applicability to long-term planning rather than real-time emergency management (Shen et al., 2019 ; Xu et al., 2024 ). Therefore, there is a lack of effective methods capable of both rapidly predicting flood inundation and concurrently providing location-specific identification of its dominant drivers. The emergence of explainable Artificial Intelligence (XAI) offers promising avenues to address the interpretability limitations inherent in 'black-box' models. By integrating XAI techniques with machine learning (ML) models, explainable machine learning models make it possible to concurrently generate predictions and quantify the contributions of drivers for specific instances (Liu et al. 2025 ). Prominent methods such as SHAP (Lundberg and Lee 2017 ) and LIME (Ribeiro et al. 2016 ) are capable of providing these explanations for individual predictions. For instance, Yang et al.(2020) utilized Shapley values to identify rainfall, snowmelt, and wetness as key drivers of peak flow in numerous Chinese catchments. Zahura and Goodall ( 2022 ) employed LIME to show that local features like elevation and hourly rainfall significantly impacted flood depth estimations, with contributions varying by flood event type. Furthermore, Jiang et al.(2024) used SHAP to quantify the relative importance of interacting drivers in diverse flood magnitudes across thousands of global basins. Despite the exploratory use of explainable machine learning models for flood inundation simulation and associated driver analysis in some case studies, a systematic evaluation is needed to determine how to best pair an XAI method with a high-performing ML model to ensure both predictive accuracy and robust interpretability. This paper systematically compares various explainable machine learning models for predicting maximum inundation extents and identifying their dominant drivers. We developed and evaluated a set of explainable machine learning models (EML). We constructed these models by pairing two representative XAI techniques, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), with three distinct types of ML models: a linear model (Logistic Regression), a tree-based ensemble (Random Forest), and a neural network (Multilayer Perceptron). The study area of Norfolk, Virginia, USA, was selected to train and test the models, using data from 26 historical compound flood events driven by heavy rainfall and high tides. This research aims to demonstrate the value of explainable machine learning methods in urban flood analysis, guiding informed decision-making and effective flood management in the face of growing compound flood risks, thereby advancing more targeted flood emergency management. 2 Study area and Data 2.1 Study area Norfolk, Virginia (USA), is a coastal city in southeastern Virginia, situated at the confluence of the Elizabeth River and Chesapeake Bay (Fig. 1 ). With an extensive 144 km shoreline, the city serves as a significant port. Norfolk's topography is exceptionally low-lying, with an average elevation of only 1.8 meters and a maximum of 12 meters. This makes it one of the U.S. coastal cities most susceptible to flooding, frequently experiencing inundation from storm surges and rising sea levels. Furthermore, the August-September period often brings increased tropical cyclone activity (hurricanes and tropical storms), resulting in heavy rainfall. The combination of these coastal and rainfall-driven factors leads to frequent compound flooding, posing a persistent challenge to the city (Zahura et al. 2020 ). Given Norfolk's vulnerability to such events and the availability of comprehensive street-scale meteorological, topographic, and water depth data ( Zahura and Goodall, 2022 ), this study focuses on an approximately 56.4 km² urbanized area in the southwestern coastal part of the city. 2.2 Data preparation 2.2.1 Deriving maximum inundation extent for road segments In the study area, flooding is primarily caused by rainfall and sea level rise. Flood events were identified through the selection of the top 16 rainfall events with the highest intensity and the top 10 tidal events with the highest sea levels, occurring from 2016 to 2018 (Table 2 , Figures A1 and A2). The water depth data utilized in this research were originally simulated using the TUFLOW model, a one-dimensional (1D) pipe network and two-dimensional (2D) overland hydrodynamic flood model. This model simulated inundation across the entire city of Norfolk for the 26 identified flood events (Zahura and Goodall 2022 ). The TUFLOW model provides water depth estimates at a 5-meter spatial resolution and a 1-hour temporal resolution. The model's calibration and validation have been previously established, ensuring the accuracy of its outputs (Shen et al., 2019 ). Based on this existing research, hourly inundation depths were extracted for individual road segments. The road network was discretized into 16,868 segments, each 50 meters in length and 7.2 meters in width. Within each segment, flood conditioning factors and inundation depths were assumed to be homogeneous. To determine the maximum inundation extents, we converted the simulated hourly water depth sequences for each road segment into binary classifications (i.e., flooded or non-flooded) using a threshold-based approach. This process involved first extracting the hourly simulated water depth data for every segment across all 26 flood events. Subsequently, the maximum water depth experienced by each segment during each event was identified. A flooding threshold of 0.3 meters, corresponding to the approximate height of a typical passenger car's exhaust outlet, was then applied to classify segments as either flooded or non-flooded. This procedure resulted in the identification of 28,650 flooded road segment instances. The maximum inundation area for each event was subsequently calculated by multiplying the total number of flooded segments by the area of an individual segment. To ensure the representativeness of our training and testing datasets, we then employed an interval sampling method to select 13 events for the training set and the remaining 13 events for the testing set (Table 1 ). Table 2 26 rainfall and high tide flood events. Flood event date Rainfall intensity averaged across all segments (mm/min) Maximum tide level (m) Train or test Number of flooded segments 8/20/2018 84.25 0.610 Train 1056 10/08/2016 80.99 0.739 Test 4130 9/21/2016 37.66 0.734 Train 2743 7/30/2018 37.47 0.479 Test 903 8/29/2017 34.53 1.094 Train 1513 9/03/2016 32.91 1.377 Test 2384 7/15/2017 32.67 0.441 Train 1684 7/31/2016 31.21 0.475 Test 2427 5/28/2018 27.72 0.633 Train 1044 8/07/2017 22.96 0.445 Test 850 6/05/2016 22.55 0.624 Train 1569 1/02/2017 21.28 0.390 Test 595 5/06/2018 20.85 0.244 Train 1619 10/29/2017 20.69 0.603 Test 649 6/22/2018 17.6 0.684 Train 1316 8/09/2016 16.49 0.401 Test 227 1/23/2016 12.11 1.114 Train 1313 9/19/2017 9.97 1.098 Test 195 9/09/2018 5.31 1.082 Train 867 5/05/2016 4.4 1.072 Test 355 3/21/2018 4.17 1.021 Train 386 9/29/2016 3.91 1.012 Test 159 2/09/2016 0.73 0.996 Train 131 1/24/2017 0.17 0.994 Test 144 11/08/2017 0.09 0.981 Train 316 1/23/2017 0.0 0.851 Test 75 2.2.2 Flood conditioning factors Nine key flood conditioning factors were selected for this study, as detailed in Table 3 . The selection was informed by a review of relevant literature, considering the specific characteristics of coastal urban environments and data availability within the study area. These factors are categorized into three groups: rainfall, tidal, and topographical. Four rainfall-related factors were incorporated to characterize precipitation patterns: 1. Maximum 1-hour rainfall (Max_1HR): This represents the highest recorded rainfall intensity within a one-hour period for each road segment, reflecting the immediate impact of intense precipitation on surface inundation. 2. Cumulative rainfall in the first 2 hours (Max_1HR_2) and 72 hours (Max_1HR_72): These metrics quantify the total rainfall accumulated within 2 and 72 hours, respectively, relative to the timing of the peak one-hour rainfall for each flood event. Max_1HR_2 and Max_1HR_72 serve as proxies for the antecedent conditions, capturing the influence of recent and preceding rainfall on surface inundation. 3. Total hours of rainfall exceeding 0 mm (RH_Count): This factor provides information on the duration of a rainfall event. Longer rainfall durations are generally associated with an increased likelihood of surface flooding. Two tidal factors were utilized: 1. Maximum 1-hour tide level (Max_1TD): This denotes the peak tide level observed within a one-hour window for each road segment during a flood event, highlighting the direct influence of tidal surges on inundation. 2. Total hours tide level exceeded 0.8 m (TD_Count): In the study area, tidal flooding typically occurs when tide levels surpass 0.8 meters (Zahura and Goodall 2022 ). TD_Count quantifies the duration for which this critical threshold is exceeded, with prolonged periods of high tides increasing the probability of surface inundation. Three topographical features were included: 1. Elevation (EV): Previous research has consistently identified elevation as a crucial factor in flood susceptibility (Chapi et al. 2017 ; Li et al. 2023 ), with lower-lying, flatter areas being inherently more prone to flooding than areas at higher elevations. 2. Distance to Water (DTW): This factor measures the shortest distance from a given road segment to the nearest surface water body. Smaller DTW values indicate closer proximity to water sources, which often correlates with wetter soils and a consequently higher likelihood of surface flood inundation (Lyu and Yin 2023 ). 3. Topographic Wetness Index (TWI): TWI is a steady-state wetness index that quantifies the potential for water to accumulate at any point in a catchment, thereby representing the propensity for water accumulation within the urban roadway network (Manfreda et al., 2014 ). Table 3 Factors used in flood inundation areas modeling for Norfolk Category No Feature/name (abbreviation) Unit Data source Rainfall 1 Maximum one-hour rainfall (Max_1HR) mm Hourly rainfall observations were obtained from Hampton Roads Sanitation District (HRSD) observation sites 2 Cumulative rainfall for the previous 2hrs (Max_1HR_2) mm 3 Cumulative rainfall for the previous 72hrs (Max_1HR_72) mm 4 Total hours of rainfall exceeding 0 mm (RH_Count) hour Tidal 5 Maximum one-hour tide level (Max_1TD) m Hourly tide levels were obtained from NOAA's Sewells Point station 6 Total number of hours that the tide level exceeds 0.8m(TD_Count) hour Topographic 7 Elevation (EV) m A 5 m Digital Elevation Model (DEM) was obtained from the U.S. Geological Survey (USGS) 8 Depth to water index (DTW) cm 9 Topographic wetness index (TWI) - 3 Methods Figure 2 illustrates the research workflow. It begins with the data preparation phase (detailed in the previous section). Subsequently, the explainable machine learning stage employs the training data to develop flood inundation predictive models, specifically Linear Regression (LR), Random Forest (RF), and Multilayer Perceptron (MLP). Explainable AI techniques were then employed to interpret the predictions generated by these models and to determine the dominant flood drivers, followed by an evaluation of the accuracy of the flood inundation predictions and the dominant driver identifications. 3.1 Maximum inundation extent prediction 3.1.1 Machine learning models In this study, we selected representative models from each type to model flood maximum inundation areas: LR (linear), RF (tree-based), and MLP (network-based). Logistic regression (LR), introduced by Cox ( 1958 ), is a linear statistical model that establishes a relationship between a set of predictor variables (i.e., conditioning factors) and a binary dependent variable. Pradhan ( 2009 ) first applied LR to flood inundation area modeling. In this context, the LR model is trained on historical data to discern the relationship between various geospatial and event-specific features and binary flood labels (e.g., flooded/non-flooded). The model employs a sigmoid function to transform a linear combination of input features into a probability estimate (ranging from 0 to 1) of flooding for each segment. Random Forest (RF), proposed by Breiman ( 2001 ), is an ensemble tree-based algorithm widely adopted for flood inundation area modeling since its initial application in this field by Trigila et al.(2015). RF constructs multiple decision trees during training and outputs the mode of the classes (for classification) or mean prediction (for regression) of the individual trees, thereby enhancing predictive accuracy and stability. In flood inundation modeling, RF offers several advantages, including robustness to multicollinearity among predictor variables, effective handling of missing data, and a significant reduction in overfitting compared to individual decision trees (Choubin et al. 2019 ). These characteristics make it particularly well-suited for complex flood prediction tasks. Additionally, neural network algorithms have been adopted in flood modelling due to their strong predictive capabilities (Zhao et al. 2018 ). Various neural network models, including MLP, CNN, RNN and other more advanced deep learning models, have been widely used in this field (Tien Bui et al. 2020 ). Among these, MLP is the fundamental algorithm, consisting of multiple layers of computational units (i.e., nodes or neurons) connected by weights. In the MLP approach, input-output relationships are established through multiple layers of nonlinear transformations. The strength of neural network models lies in their ability to address complex nonlinear problems. 3.1.2 Model training The training process for the models involves hyperparameter optimization and implementation. Each model has several key hyperparameters that greatly influence performance, making tuning a crucial step. Hyperparameter tuning was conducted via 5-fold cross-validation using "GridSearchCV" for automatic optimization. Given the focus on accurate flood inundation detection, recall was used as the criterion for optimization. Table 4 provides the optimal hyperparameters for all models. The construction and training of LR, RF, and MLP models were implemented using the LogisticRegression, RandomForestRegressor, and MLPClassifier modules of the scikit-learn library in Python. Table 4 Description of models and their hyperparameters. Class Model Hyper-parameters Range Optimized value Linear LR penalty c solver max_iter [‘L1’, ‘L2’] [0.1, 1, 10] [‘liblinear’, ‘saga’] [100, 200, 500] L1 10 liblinear 100 Tree RF n_estimators max_depth min_samples_split min_samples_leaf [50, 100, 200] [5, 10, 20, None] [2, 5, 10] [1, 2, 4] 150 None 5 2 ccp_alpha [0.0002, 0.00025, 0.0003] 0.00025 Neural network MLP hidden_layer_sizes activation solver learning_rate max_iter [(10,), (50,), (100,)] [‘relu’, ‘tanh’] [‘sgd’, ‘adam’] [‘constant’, ‘adaptive’] [100, 200, 500] (100,) tanh adam constant 500 3.1.3 Evaluation metrics The performance of the machine learning models was evaluated using several standard metrics, including the Receiver Operating Characteristic (ROC) curve, the Area Under the Curve (AUC), accuracy, precision, recall, and the F1 score, as shown in Table 5 . These metrics are derived from the classification outcomes for each road segment, which can be categorized into four types based on a confusion matrix. True Positive (TP): Represents the number of road segments correctly predicted as flooded. False Positive (FP): Represents the number of non-flooded road segments incorrectly predicted as flooded. True Negative (TN): Represents the number of non-flooded road segments correctly predicted as non-flooded. False Negative (FN): Represents the number of flooded road segments incorrectly predicted as non-flooded. Table 5 Model Performance Evaluation Metrics Evaluation criterion Description & formula Ture positive (TP) Number of flood samples correctly classified. False Positive (FP) Number of flood samples not correctly classified. True Negative (TN) Number of non-flood samples correctly classified. False Negative (FN) Number of non-flood samples not correctly classified. Accuracy (Acc) \(\:Accuracy=\:\frac{TP+TN}{TP+TN+FP+FN}\) Recall (R) \(\:Recall=\:\frac{TP}{TP+FN}\) Precision (P) \(\:Precision=\:\frac{TP}{TP+FP}\) F1 score (F) \(\:F1\:score=2\times\:\frac{P\times\:R}{P+R}\) 3.2 Dominant flood driver identification 3.2.1 Explainable AI (XAI) Techniques To identify the dominant drivers of flood inundation, this study employed two widely recognized, model-agnostic explainable AI (XAI) techniques: Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), as shown in Fig. 3 . LIME, developed by Ribeiro et al.(2016), is designed to explain the predictions of any machine learning model for individual instances. It operates by approximating the behavior of a complex model locally with a simpler, more interpretable model (e.g., a linear model). The core assumption is that while the global behavior of a complex model may be intricate, its decision-making process for a specific instance can be faithfully represented by a simpler model learned in the vicinity of that instance. Operationally, LIME generates a new dataset of perturbed samples around the instance of interest, obtains predictions for these samples from the original complex model, and then trains an interpretable model on this new dataset weighted by proximity to the original instance. The parameters of this local interpretable model (such as the weights in a linear regression) then serve as explanations for the individual prediction, indicating the contribution of each feature to that specific outcome. This approach enhances the interpretability of complex models such as random forests and neural networks. SHAP, introduced by Lundberg and Lee ( 2017 ), is another prominent technique for explaining model predictions. It leverages Shapley values, a concept from cooperative game theory, to attribute the prediction outcome to each feature. A feature's Shapley value represents its average marginal contribution to the prediction across all possible combinations of features, thereby capturing both its individual impact and its interactions with other features. For a given prediction, the sum of the SHAP values for all features, plus a baseline value (typically the average prediction over the training dataset), equals the model's actual output for that instance. SHAP values thus provide a consistent and locally accurate measure of feature importance, ensuring that the explanation is fair and reliable. 3.2.2 Identifying dominant flood drivers using XAI techniques The process for identifying dominant flood drivers using the selected XAI techniques involved the following steps: 1. Feature contribution calculation: Subsequent to the training of the machine learning models, both LIME and SHAP were applied to interpret the model predictions. This involved calculating the specific contribution of each input feature to the predicted flood status for every individual road segment within each flood event. 2. Extraction of driver-specific contributions: From the comprehensive set of feature contributions, values corresponding to the four designated rainfall features and the two designated tidal features were specifically extracted for further analysis. 3. Dominant driver classification: For each road segment predicted as inundated, the feature (from the aforementioned rainfall and tidal categories) with the highest positive contribution to the flood prediction was identified. Based on whether this most influential feature was rainfall-related or tidal-related, the inundated road segment was classified as either rainfall-dominated or tidal-dominated. 4 Results 4.1 Predictive performance of ML models The predictive performance of machine learning models for maximum flood inundation was assessed using Accuracy, Precision, Recall, F1 score, and ROC-AUC, with results summarized in Table 6 . RF model consistently outperformed MLP and LR. On the testing set, RF achieved the highest Accuracy (0.79, compared to MLP's 0.75 and LR's 0.74) and Recall (0.81, vs. MLP 0.78, LR 0.67), demonstrating its effectiveness in minimizing the omission of true flood events. For Precision, RF and LR (both 0.78) surpassed MLP (0.74), indicating reliable positive predictions when a segment was classified as flooded. RF also led with the top F1 score of 0.79 (MLP 0.76, LR 0.72). Collectively, these results underscore RF's superior robustness and reliability in predicting maximum flood inundation, establishing it as the most effective model among those evaluated. Table 6 Predictive performance evaluation of each ML model. Training Testing LR RF MLP LR RF MLP Accuracy 0.76 0.84 0.78 0.74 0.79 0.75 Precision 0.78 0.80 0.77 0.78 0.78 0.74 Recall 0.73 0.89 0.78 0.67 0.81 0.78 F1 score 0.75 0.84 0.78 0.72 0.79 0.76 ROC curve analysis (Fig. 4 ) further highlighted the RF model's superior performance. On the test dataset, RF achieved the highest AUC with a value of 0.870, outperforming both MLP (AUC 0.829) and LR (AUC 0.809). Considering its consistently strong performance across all evaluation metrics, the RF model was therefore selected for the subsequent identification of dominant flood drivers. 4.2 Predicted maximum flood inundation maps Given the extensive number of flood events in the test set, three representative events were selected for a detailed spatial analysis of RF model's predictions. These events were chosen based on differing rainfall intensity and tidal levels: (1) A high rainfall intensity, low tidal level event (July 31, 2016), (2) A low rainfall intensity, high tidal level event (January 24, 2017), (3) A high rainfall intensity, high tidal level event (September 3, 2016). Figure 5 illustrates the spatial distribution of the predicted maximum flood inundation for these events, with prediction outcomes categorized by True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). For the high rainfall, low tidal level event (Fig. 5 a-b), the RF model correctly identified 1,874 out of 2,427 actual inundated road segments, achieving a recall of approximately 77%. The inundated segments were widely distributed, with notable concentrations in the southwestern part of the study area, particularly in downtown Norfolk. In the low rainfall, high tidal level event (Fig. 5 c-d), the model demonstrated high performance, correctly predicting 140 out of 144 actual inundated segments. This resulted in a recall of approximately 97%. Unlike the widespread inundation in the high-rainfall event, the flooded segments in this scenario were more sparsely distributed and primarily located near the central coastal area. For the high rainfall, high tidal level event (Fig. 5 e-f), the model correctly identified 1,806 out of 2,384 actual inundated segments, corresponding to a recall of approximately 76%. The correctly predicted segments were extensively distributed across both coastal and inland areas. This widespread inundation likely resulted from the combined influence of significant rainfall and high tidal levels, which jointly increased the probability of flooding. Overall, the visual and quantitative analysis of these representative events (Fig. 5 ) demonstrates the RF model's robust capability in predicting the spatial extent of flood inundation under varied environmental conditions. 4.3 Dominant flood driver analysis: Comparing SHAP and LIME The SHAP and LIME techniques were utilized to analyze the contributions of various rainfall and tidal features to the flood predictions made by the RF model, specifically for those road segments correctly identified as inundated. The dominant driver (rainfall or tidal) for each such segment was determined by identifying whether the feature exhibiting the highest positive contribution was rainfall-related or tidal-related. Given the lack of direct ground-truth observations detailing the actual dominant flood driver for each road segment during each historical event, this study assesses the spatial reasonableness of the derived driver distributions to compare the effectiveness of the different methods. The evaluation is based on the premise that a plausible spatial pattern would generally conform to expected hydrological behaviors in coastal cities. Specifically, it is assumed that rainfall-dominated inundation would likely be more dispersed throughout the study area, whereas tidal-dominated inundation would primarily be concentrated along the coastline and tidally influenced water bodies. The consistency of the XAI-derived driver patterns with these expected distributions serves as an indicator of their rationality. The comparative results for three representative flood events are presented in Fig. 6 . During the January 24, 2017 event (Fig. 6 b, e), characterized by high tidal levels and low rainfall, SHAP and LIME again demonstrated close agreement. Both methods predominantly classified the correctly predicted flooded segments as tide-dominated (SHAP: 136/140, ~ 97.1%; LIME: 138/140, ~ 98.6%). For both XAI tools, TD_Count was overwhelmingly the principal feature identified for these tide-dominated instances. As anticipated, these segments were primarily concentrated along the coast, consistent with tidal influence patterns. Significant difference between SHAP and LIME occurred for the September 3, 2016 compound event (high rainfall and high tidal levels; Fig. 6 c,f), which involved 1,806 correctly predicted flooded segments. SHAP classified 1,168 segments (approx. 64.7%) as rainfall-dominated and 638 segments (approx. 35.3%) as tide-dominated. Among SHAP's rainfall-dominated segments, Max_1HR_72 was the leading contributor (approx. 51%), followed by RH_Count (approx. 8%) and Max_1HR_2 (approx. 5%). For SHAP's tide-dominated segments, TD_Count was the dominant feature in approximately 96% of cases. LIME classified approximately 1,427 segments (79% of 1,806) as rainfall-dominated and approximately 379 segments (21% of 1,806) as tide-dominated. LIME attributed about 68% of its rainfall-dominated segments to Max_1HR_72, with TD_Count and RH_Count also noted as significant contributors. For LIME's tide-dominated segments, TD_Count reportedly influenced 66% and Max_1TD influenced 34%. Spatially, SHAP’s results for the compound event (Fig. 6 c) depicted tide-dominated segments clustered near the coast and rainfall-dominated segments dispersed throughout the study area, which is consistent with expected hydrological patterns. In contrast, LIME’s results (Fig. 6 f) showed a less intuitive distribution, placing many tide-dominated segments inland (particularly in the northern region) and concentrating rainfall-dominated segments along the coast, thereby deviating from empirical patterns and the physical understanding of these drivers. Overall, the results obtained from SHAP demonstrate a closer alignment with expected flood patterns in coastal cities. 5 Discussions 5.1 Effectiveness of explainable ML models This study evaluated three ML models—LR, RF, and MLP—along with two XAI techniques, LIME and SHAP. The evaluation centered on two key aspects: the accuracy of flood inundation area prediction and the spatial reasonableness of dominant flood driver identification. Among the ML model-XAI combinations assessed, the RF model paired with SHAP yielded the most robust performance. Regarding predictive capability for flood inundation, the tree-structured RF model surpassed the other models, achieving an accuracy of 0.79 in discriminating between flooded and non-flooded road segments. This superior performance of RF is consistent with findings from similar comparative studies in flood modeling (Rezvani et al. 2024 ). Conversely, the LR model exhibited the weakest predictive power. This outcome also aligns with its frequent use as a benchmark in existing literature, where it often serves as a baseline for comparison against more sophisticated models (Hong et al. 2018 ; Al-Juaidi 2023 ). From the perspective of identifying dominant flood drivers, this study employed XAI techniques to identify dominant drivers at specific inundation locations. This approach contrasts with many previous studies that typically offer global feature importance rankings applicable to the entire study area. XAI techniques delivers more granular, localized insights into flood causality (Avand et al. 2021 ; Mahdizadeh Gharakhanlou and Perez 2023 ), which can be more actionable for targeted flood mitigation. 5.2 Rationale for RF-SHAP's effectiveness The study's results demonstrate that RF combined with SHAP achieved superior performance in both predicting flood inundation areas (Table 6 ) and identifying dominant flood drivers (Fig. 6 ). The reasons for this are as follows. RF's effectiveness in flood extent prediction lies in its ensemble structure, which robustly captures the complex non-linear relationships inherent in flood phenomena, often with less extensive data and tuning requirements than models like MLP (Tyralis et al. 2019 ). For dominant flood driver identification, SHAP demonstrated advantages over LIME, particularly in accurately representing driver influence during complex compound flood events (e.g., Fig. 6 c,f). The primary reason is SHAP's methodology, which inherently accounts for feature interactions by evaluating how the contribution of one feature is affected by the presence or absence of others (Lundberg and Lee 2017 ). This is crucial because flood drivers (like rainfall and tidal influences) are often interdependent. LIME, by typically relying on local linear approximations and assuming feature independence, is less adept at capturing these critical interaction effects (Ribeiro et al. 2016 ). Given that the interplay between different factors significantly influences flood occurrence and characteristics, SHAP’s ability to incorporate these interactions into its attributions provides a more reliable and mechanistically sound identification of dominant flood drivers compared to LIME in such complex environmental systems. 5.3 Leveraging explainable ML model for enhanced flood emergency response Explainable machine learning models like RF-SHAP, validated in this study, offer significant practical value for flood emergency response. By identifying location-specific dominant flood drivers (e.g., tidal vs. rainfall), the method helps authorities develop more targeted interventions and optimize resource allocation (Cools et al. 2016 ). It also overcomes the "black box" problem of traditional AI by explaining the reasoning behind its predictions, which builds trust and leads to more confident, science-based decisions by managers (Adadi and Berrada 2018 ). Crucially, the model is highly efficient, providing near real-time prediction and attribution (approx. 33 seconds per event), which is essential for managing rapidly changing flood conditions. 5.4 Limitations and future work Despite the demonstrated effectiveness of the RF-SHAP approach for flood analysis in Norfolk, Virginia, this study has limitations that suggest avenues for future research. A primary limitation is that the methodology's validation is currently geographically specific. Future work should therefore evaluate its performance and adaptability in diverse regions with varying hydrogeological and urban characteristics to ascertain broader generalizability. Furthermore, this study focused on maximum flood inundation, not the dynamic spatiotemporal evolution of flooding. Future research could aim to develop models that dynamically predict flood progression over time and space, concurrently identifying how dominant drivers evolve throughout a flood event. This would offer a more comprehensive understanding, particularly valuable for real-time risk assessment and response. 6 Conclusions For effective emergency response in urban areas vulnerable to compound flooding, the rapid identification of inundation areas and their dominant drivers is essential. This study addressed this dual challenge by systematically evaluating explainable machine learning (EML) frameworks that pair machine learning (ML) models with explainable AI (XAI) techniques. Our principal findings show that the combination of Random Forest (RF) and SHAP is the most effective approach for both rapid flood prediction and reliable driver identification. The superiority of this pairing stems from SHAP's ability to accurately quantify driver influence, even in complex compound events, by accounting for the critical interactions between flood factors—a key limitation of other methods like LIME, which largely assume feature independence. The integration of the RF-SHAP model into flood management offers a powerful tool to substantially enhance emergency response. By providing rapid, spatially-specific inundation predictions and interpretable insights into dominant drivers, this approach enables more targeted strategies, optimized resource allocation, and confident decision support. These advancements are crucial for building resilience in coastal cities facing growing compound flood risks. Declarations Acknowledgements This work was supported by the National Natural Science Foundation of China (Grant No. 52309005), the Fundamental Research Funds for the Central Universities (Grant No. DUT24ZD409), the National Natural Science Foundation of China (Grant No. U2240204), and the National Key R&D Program of China (Grant No. 2024YFC3213000). We would also like to thank Dr. Faria Zahura and her team for providing the foundational data that greatly facilitated our research. Funding This work was supported by the National Natural Science Foundation of China (Grant No. 52309005), the Fundamental Research Funds for the Central Universities (Grant No. DUT24ZD409), the National Natural Science Foundation of China (Grant No. U2240204), and the National Key R&D Program of China (Grant No. 2024YFC3213000). Competing Interests The authors have no relevant financial or non-financial interests to disclose. Author Contributions All authors contributed to the study conception and design. Jiqiang Xie: Writing – original draft, Methodology, Data curation, Conceptualization. Heng Lyu: Writing – review & editing, Funding acquisition, Conceptualization. Bing Yu: Review & editing, Conceptualization. Shengnan Fu: Writing – review & editing. Chen Yang: Methodology, Conceptualization. Chi Zhang: Supervision. References Adadi A, Berrada M (2018) Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access 6:52138–52160. https://doi.org/10.1109/ACCESS.2018.2870052 Al-Juaidi AEM (2023) The interaction of topographic slope with various geo-environmental flood-causing factors on flood prediction and susceptibility mapping. Environ Sci Pollut Res 30:59327–59348. https://doi.org/10.1007/s11356-023-26616-y Asif M, Kuglitsch MM, Pelivan I, Albano R (2025) Review and Intercomparison of Machine Learning Applications for Short-term Flood Forecasting. Water Resour Manag 39:1971–1991. https://doi.org/10.1007/s11269-025-04093-x Avand M, Moradi H, lasboyee MR (2021) Spatial modeling of flood probability using geo-environmental variables and machine learning models, case study: Tajan watershed, Iran. Adv Space Res 67:3169–3186. https://doi.org/10.1016/j.asr.2021.02.011 Bentivoglio R, Isufi E, Jonkman SN, Taormina R (2022) Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrol Earth Syst Sci 26:4345–4378. https://doi.org/10.5194/hess-26-4345-2022 Breiman L (2001) Random Forests. Mach Learn 45:5–32. https://doi.org/10.1023/A:1010933404324 Chapi K, Singh VP, Shirzadi A, et al (2017) A novel hybrid artificial intelligence approach for flood susceptibility assessment. Environ Model Softw 95:229–245. https://doi.org/10.1016/j.envsoft.2017.06.012 Choubin B, Moradi E, Golshan M, et al (2019) An ensemble prediction of flood susceptibility using multivariate discriminant analysis, classification and regression trees, and support vector machines. Sci Total Environ 651:2087–2096. https://doi.org/10.1016/j.scitotenv.2018.10.064 Cools J, Innocenti D, O’Brien S (2016) Lessons from flood early warning systems. Environ Sci Policy 58:117–122. https://doi.org/10.1016/j.envsci.2016.01.006 Cox DR (1958) The Regression Analysis of Binary Sequences. J R Stat Soc Ser B Methodol 20:215–232. https://doi.org/10.1111/j.2517-6161.1958.tb00292.x El baida M, Boushaba F, Chourak M, Hosni M (2024) Real-Time Urban Flood Depth Mapping: Convolutional Neural Networks for Pluvial and Fluvial Flood Emulation. Water Resour Manag 38:4763–4782. https://doi.org/10.1007/s11269-024-03886-w Fang J, Wahl T, Fang J, et al (2021) Compound flood potential from storm surge and heavy precipitation in coastal China: dependence, drivers, and impacts. Hydrol Earth Syst Sci 25:4403–4416. https://doi.org/10.5194/hess-25-4403-2021 Han S, Tahvildari N (2024) Compound Flooding Hazards Due To Storm Surge and Pluvial Flow in a Low-Gradient Coastal Region. Water Resour Res 60:e2023WR037014. https://doi.org/10.1029/2023WR037014 Hendry A, Haigh ID, Nicholls RJ, et al (2019) Assessing the characteristics and drivers of compound flooding events around the UK coast. Hydrol Earth Syst Sci 23:3117–3139. https://doi.org/10.5194/hess-23-3117-2019 Hong H, Tsangaratos P, Ilia I, et al (2018) Application of fuzzy weight of evidence and data mining techniques in construction of flood susceptibility map of Poyang County, China. Sci Total Environ 625:575–588. https://doi.org/10.1016/j.scitotenv.2017.12.256 Jiang S, Tarasova L, Yu G, Zscheischler J (2024) Compounding effects in flood drivers challenge estimates of extreme river floods. Sci Adv 10:eadl4005. https://doi.org/10.1126/sciadv.adl4005 Kabir S, Patidar S, Xia X, et al (2020) A deep convolutional neural network model for rapid prediction of fluvial flood inundation. J Hydrol 590:125481. https://doi.org/10.1016/j.jhydrol.2020.125481 Li Y, Osei FB, Hu T, Stein A (2023) Urban flood susceptibility mapping based on social media data in Chengdu city, China. Sustain Cities Soc 88:104307. https://doi.org/10.1016/j.scs.2022.104307 Liu Y, Wang H, Guan X, et al (2025) Urban Flood Depth Prediction and Visualization Based on the XGBoost-SHAP Model. Water Resour Manag 39:1353–1375. https://doi.org/10.1007/s11269-024-04020-6 Lundberg SM, Lee S-I (2017) A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Curran Associates Inc., Red Hook, NY, USA, pp 4768–4777 Lyu H-M, Yin Z-Y (2023) Flood susceptibility prediction using tree-based machine learning models in the GBA. Sustain Cities Soc 97:104744. https://doi.org/10.1016/j.scs.2023.104744 Mahdizadeh Gharakhanlou N, Perez L (2023) Flood susceptible prediction through the use of geospatial variables and machine learning methods. J Hydrol 617:129121. https://doi.org/10.1016/j.jhydrol.2023.129121 Manfreda S, Nardi F, Samela C, et al (2014) Investigation on the use of geomorphic approaches for the delineation of flood prone areas. J Hydrol 517:863–876. https://doi.org/10.1016/j.jhydrol.2014.06.009 Mangukiya NK, Kushwaha S, Sharma A (2024) A novel multi-model ensemble framework for fluvial flood inundation mapping. Environ Model Softw 180:106163. https://doi.org/10.1016/j.envsoft.2024.106163 Materia S, García LP, van Straaten C, et al (2024) Artificial intelligence for climate prediction of extremes: State of the art, challenges, and future perspectives. WIREs Clim Change 15:e914. https://doi.org/10.1002/wcc.914 Pradhan B (2009) Flood susceptible mapping and risk area delineation using logistic regression, GIS and remote sensing. J Spat Hydrol 9:1–18 Qin H, Liang Q, Chen H, De Silva V (2024a) A Coupled Human and Natural Systems (CHANS) framework integrated with reinforcement learning for urban flood mitigation. J Hydrol 643:131918. https://doi.org/10.1016/j.jhydrol.2024.131918 Qin H, Liang Q, Chen H, De Silva V (2024b) A two-way coupled CHANS model for flood emergency management, with a focus on temporary flood defences. Environ Model Softw 181:106166. https://doi.org/10.1016/j.envsoft.2024.106166 Razavi S, Tolson BA, Burn DH (2012) Review of surrogate modeling in water resources. Water Resour Res 48:. https://doi.org/10.1029/2011WR011527 Rezvani SM, Gonçalves A, Silva MJF, de Almeida NM (2024) Smart hotspot detection using geospatial artificial intelligence: A machine learning approach to reduce flood risk. Sustain Cities Soc 115:105873. https://doi.org/10.1016/j.scs.2024.105873 Ribeiro MT, Singh S, Guestrin C (2016) “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, New York, NY, USA, pp 1135–1144 Rohmer J, Idier D, Paris F, et al (2018) Casting light on forcing and breaching scenarios that lead to marine inundation: Combining numerical simulations with a random-forest classification approach. Environ Model Softw 104:64–80. https://doi.org/10.1016/j.envsoft.2018.03.003 Rong Y, Bates P, Neal J (2024) GPU‐Accelerated Urban Flood Modeling Using a Nonuniform Structured Grid and a Super Grid Scale River Channel. Water Resour Res e2023WR036128: https://doi.org/10.1029/2023WR036128 Rudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell 1:206–215. https://doi.org/10.1038/s42256-019-0048-x Sanders BF, Schubert JE (2019) PRIMo: Parallel raster inundation model. Adv Water Resour 126:79–95. https://doi.org/10.1016/j.advwatres.2019.02.007 Shen Y, Morsy MM, Huxley C, et al (2019) Flood risk assessment and increased resilience for coastal urban watersheds under the combined impact of storm tide and heavy rainfall. J Hydrol 579:124159. https://doi.org/10.1016/j.jhydrol.2019.124159 Sun J, Chen W, Hu B, et al (2023) Roles of reservoirs in regulating basin flood and droughts risks under climate change: Historical assessment and future projection. J Hydrol Reg Stud 48:101453. https://doi.org/10.1016/j.ejrh.2023.101453 Teng J, Jakeman AJ, Vaze J, et al (2017) Flood inundation modelling: A review of methods, recent advances and uncertainty analysis. Environ Model Softw 90:201–216. https://doi.org/10.1016/j.envsoft.2017.01.006 Tian Y, Cui R, Fu W, et al (2025) Dual-modality Encoder-decoder Framework for Urban Real-time Rainfall-runoff Prediction. Water Resour Manag 39:4509–4538. https://doi.org/10.1007/s11269-025-04166-x Tien Bui D, Hoang N-D, Martínez-Álvarez F, et al (2020) A novel deep learning neural network approach for predicting flash flood susceptibility: A case study at a high frequency tropical storm area. Sci Total Environ 701:134413. https://doi.org/10.1016/j.scitotenv.2019.134413 Trigila A, Iadanza C, Esposito C, Scarascia-Mugnozza G (2015) Comparison of Logistic Regression and Random Forests techniques for shallow landslide susceptibility assessment in Giampilieri (NE Sicily, Italy). Geomorphology 249:119–136. https://doi.org/10.1016/j.geomorph.2015.06.001 Tyralis H, Papacharalampous G, Langousis A (2019) A Brief Review of Random Forests for Water Scientists and Practitioners and Their Recent History in Water Resources. Water 11:910. https://doi.org/10.3390/w11050910 Wahl T, Jain S, Bender J, et al (2015) Increasing risk of compound flooding from storm surge and rainfall for major US cities. Nat Clim Change 5:1093–1097. https://doi.org/10.1038/nclimate2736 Xu H, Ragno E, Jonkman SN, et al (2024) Combining statistical and hydrodynamic models to assess compound flood hazards from rainfall and storm surge: a case study of Shanghai. Hydrol Earth Syst Sci 28:3919–3930. https://doi.org/10.5194/hess-28-3919-2024 Yang W, Yang H, Yang D (2020) Classifying floods by quantifying driver contributions in the Eastern Monsoon Region of China. J Hydrol 585:124767. https://doi.org/10.1016/j.jhydrol.2020.124767 Yin J, Yang Y, Yu D, et al (2024) Strategic storm flood evacuation planning for large coastal cities enables more effective transfer of elderly populations. Nat Water 2:274–284. https://doi.org/10.1038/s44221-024-00210-z Zahura FT, Goodall JL (2022) Predicting combined tidal and pluvial flood inundation using a machine learning surrogate model. J Hydrol Reg Stud 41:101087. https://doi.org/10.1016/j.ejrh.2022.101087 Zahura FT, Goodall JL, Sadler JM, et al (2020) Training Machine Learning Surrogate Models From a High-Fidelity Physics-Based Model: Application for Real-Time Street-Scale Flood Prediction in an Urban Coastal Community. Water Resour Res 56:e2019WR027038. https://doi.org/10.1029/2019WR027038 Zhao G, Pang B, Xu Z, et al (2018) Mapping flood susceptibility in mountainous areas on a national scale in China. Sci Total Environ 615:1133–1142. https://doi.org/10.1016/j.scitotenv.2017.10.037 Zhao M, Liu Y, Wang Y, et al (2022) Effectiveness assessment of reservoir projects for flash flood control, water supply and irrigation in Wangmo Basin, China. Sci Total Environ 851:157918. https://doi.org/10.1016/j.scitotenv.2022.157918 Zhou S, Jia W, Wang M, et al (2024) Synergistic assessment of multi-scenario urban waterlogging through data-driven decoupling analysis in high-density urban areas: A case study in Shenzhen, China. J Environ Manage 369:122330. https://doi.org/10.1016/j.jenvman.2024.122330 Cite Share Download PDF Status: Under Review Version 1 posted Editor invited by journal 12 Mar, 2026 Reviewers agreed at journal 28 Sep, 2025 Reviewers invited by journal 28 Sep, 2025 Editor assigned by journal 28 Aug, 2025 First submitted to journal 27 Aug, 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-7476416","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":521684880,"identity":"adc2f18a-e6bb-42f6-8121-e5b8c5bde980","order_by":0,"name":"Jiqiang Xie","email":"","orcid":"","institution":"Dalian University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Jiqiang","middleName":"","lastName":"Xie","suffix":""},{"id":521684881,"identity":"fb367048-91f5-4889-b629-8f43bae0f4d6","order_by":1,"name":"Bing Yu","email":"","orcid":"","institution":"Dalian University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Bing","middleName":"","lastName":"Yu","suffix":""},{"id":521684882,"identity":"4d2aace2-ca14-413c-a3d8-98081918bcdc","order_by":2,"name":"Heng Lyu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYBACPgYGNhAtx3AczGcmrIUNqsWY4TADYwNJWhIbiNcikWP24OOO2vS+wzzmDxgqrBMb2M8eIKTF3HDmmeO5Mw/zGDYwnElPbODJSyBoizRv27HcDSAtjG2HExskeAwIa/nbdizdAKzlH7FaGNtqEiBaGojRwvOsTLK37YDhzMNshTMSjqUbt/Hk4NfCz568TeJnW5083/HmDR8+1FjL9rOfwa+FQSABRB6GcEBsNvzqQdYcAJF1BNWNglEwCkbBCAYATDVCdjDWyekAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-7794-6764","institution":"Dalian University of Technology","correspondingAuthor":true,"prefix":"","firstName":"Heng","middleName":"","lastName":"Lyu","suffix":""},{"id":521684883,"identity":"89f46f11-8ad9-43a1-9afa-107153741702","order_by":3,"name":"Shengnan Fu","email":"","orcid":"","institution":"Dalian University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Shengnan","middleName":"","lastName":"Fu","suffix":""},{"id":521684884,"identity":"ea10fcf2-094e-4fdd-98b8-62e478d100b4","order_by":4,"name":"Chen Yang","email":"","orcid":"","institution":"North China University of Water Resources and Electric Power","correspondingAuthor":false,"prefix":"","firstName":"Chen","middleName":"","lastName":"Yang","suffix":""},{"id":521684885,"identity":"dfa58a2e-cadc-4c24-baa1-72230e8e8045","order_by":5,"name":"Chi Zhang","email":"","orcid":"","institution":"Dalian University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Chi","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-08-28 05:21:22","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7476416/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7476416/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93154150,"identity":"cae2940c-6dc0-49e5-ba94-cb1026d51404","added_by":"auto","created_at":"2025-10-09 15:13:05","extension":"xml","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10563,"visible":true,"origin":"","legend":"","description":"","filename":"warmWARMD2502550.xml","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/5328444459cf6384d49e0801.xml"},{"id":93154149,"identity":"c687f769-0621-4505-ae0e-6562eaff927c","added_by":"auto","created_at":"2025-10-09 15:13:05","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":936,"visible":true,"origin":"","legend":"","description":"","filename":"WARMD250255043753.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/604445d193061886c7769c15.xml"},{"id":93154175,"identity":"dc3e970a-f6b2-4dfa-b93c-f2602f82ab6b","added_by":"auto","created_at":"2025-10-09 15:13:07","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":818,"visible":true,"origin":"","legend":"","description":"","filename":"WARMD2502550Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/69637ebbba73cfc0958c7d0a.xml"},{"id":93154154,"identity":"4f7bd20e-39c2-45a1-bdff-244a740c127e","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149464,"visible":true,"origin":"","legend":"","description":"","filename":"WARMD25025500enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/b7ba388a83d6a43e2874198f.xml"},{"id":93154157,"identity":"98fead5a-56bc-46d5-8789-92387703ea1c","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1407901,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/300b03a53ea73c17f067facf.png"},{"id":93155572,"identity":"5fa11e6d-6e52-4f9f-834f-395eb1555f44","added_by":"auto","created_at":"2025-10-09 15:29:06","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":499155,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/c11b8bb81370fea8c6ed5b14.png"},{"id":93154504,"identity":"f022380b-ce6d-4d59-9997-26b5c5a6caa9","added_by":"auto","created_at":"2025-10-09 15:21:06","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":283479,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/56577fb8ba3409cf05b6eb7a.png"},{"id":93154163,"identity":"5bf31e56-ea4b-45ec-989b-90ffdc1d1cb2","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":75933,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/7d95dec808f98a2c1fd1e2d4.png"},{"id":93154159,"identity":"b6376a4d-7e19-48d0-b567-f52c30a43685","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6104797,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/3118fa150ea687db08059ed2.png"},{"id":93154162,"identity":"8838270e-6b62-4e48-a4f6-eda093e9298d","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3216925,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/5085b7275a555d7e63b95e50.png"},{"id":93154505,"identity":"7d069cea-3caa-45df-92bd-078c5643f402","added_by":"auto","created_at":"2025-10-09 15:21:06","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":179527,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/e86f282c9ff7b3e9981e36d0.png"},{"id":93154153,"identity":"d6040ba4-2962-479c-ac3a-3723c48568fb","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":134815,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/f89351fdfd4c6d8088e06075.png"},{"id":93154166,"identity":"f2e98800-8240-49f5-8265-4343fb9f133f","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62821,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/38f464754bfc3588868d90ec.png"},{"id":93154168,"identity":"c553dfac-6f31-46ef-99f1-e5d6d40fbba7","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15700,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/87345cc39b8438ffe136ba12.png"},{"id":93154161,"identity":"85ef1062-0a62-4095-aa91-c3731347264d","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":724844,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/b25944d79d1bbe0f6fa04a3f.png"},{"id":93154169,"identity":"2e814183-d04e-4f61-8bdc-3edd2b4bf8b1","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":368219,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/44e51f9beda044daf34b4eb1.png"},{"id":93154158,"identity":"d22afbd5-d5aa-4235-939c-efd3bba2eb9e","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":148576,"visible":true,"origin":"","legend":"","description":"","filename":"WARMD25025500structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/4cd0e9d496ef162e8988b89b.xml"},{"id":93154170,"identity":"c4737736-3633-4e55-91ad-954a30e976fd","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"html","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":155899,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/8c9e9dbc809c9964f324f6ae.html"},{"id":93154174,"identity":"de69235e-39b1-4a70-ad61-5e6f0e8ba77e","added_by":"auto","created_at":"2025-10-09 15:13:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":460905,"visible":true,"origin":"","legend":"\u003cp\u003eStudy area in Norfolk, VA, USA.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/84c3f22a8e6bd12e94581ede.png"},{"id":93154152,"identity":"a50bec1d-e2b5-4a96-bbe9-2e1f1c09e8f0","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":677485,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow for identifying flood inundation areas and their dominant drivers using explainable machine learning.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/058ed25602a56a118b4c9fa7.png"},{"id":93154503,"identity":"ee5f8428-3030-4969-b8af-6f3947b236eb","added_by":"auto","created_at":"2025-10-09 15:21:05","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":331165,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagram of ML model explainable methods SHAP and LIME (Adapted from Ribeiro et al., 2016; Lundberg and Lee, 2017) .\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/e77e3de75de66df3deecf3b5.png"},{"id":93154173,"identity":"17337fb5-bce2-4686-98d0-6ddfb279dc3a","added_by":"auto","created_at":"2025-10-09 15:13:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":64327,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves along with the AUC values of each ML model.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/8c7bb3116414448885011210.png"},{"id":93154156,"identity":"7c4237d8-ece9-479c-834d-18b31a7c0d8a","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":2805351,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution of RF prediction for road segments across three representative test events: (a-b) a high rainfall intensity, low tidal level event (Jul/31/2016); (c-d) a low rainfall intensity, high tidal level event (Jan/24/2017); and (e-f) a high rainfall intensity, high tidal level event (Sep/03/2016). Each point represents a road segment. Prediction outcome definitions: True Positive (TP): Road segment correctly predicted as flooded; False Positive (FP): Non-flooded road segment incorrectly predicted as flooded; True Negative (TN): Non-flooded road segment correctly predicted as not flooded; False Negative (FN): Flooded road segment incorrectly predicted as not flooded.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/090e3b6496ed312592e390a9.png"},{"id":93154160,"identity":"36eea17d-b835-4f1b-bd38-06ce5b105280","added_by":"auto","created_at":"2025-10-09 15:13:06","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1668826,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution of dominant flood drivers (rainfall-dominated or tidal-dominated) in predicted inundation areas, comparing results from RF-SHAP (a-c) and RF-LIME (d-f). The comparison covers three typical flood events: a high rainfall, low tidal level event (Jul/31/2016; a,d); a low rainfall, high tidal level event (Jan/24/2017; b,e); and a high rainfall, high tidal level event (Sep/03/2016; c,f). The results indicate that flood inundation drivers identified by SHAP generally align more closely with actual flood patterns in coastal cities, where tidal-dominated segments are typically concentrated near the coast and rainfall-dominated segments are dispersed throughout the study area.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/1fe4a39df9a848333c5f7d77.png"},{"id":93156034,"identity":"afceb40a-a72e-4e6e-a89c-8bc470728b65","added_by":"auto","created_at":"2025-10-09 15:37:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7164286,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7476416/v1/f7c0bfa0-4f34-4b0f-ac28-becf41190a1c.pdf"}],"financialInterests":"","formattedTitle":"Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning","fulltext":[{"header":"Highlights","content":"\u003col\u003e\n \u003cli\u003eA framework based on explainable machine learning models is proposed to predict flood maximum inundation extent and identify the dominant flood drivers.\u003c/li\u003e\n \u003cli\u003eSystematic comparative analysis revealed that RF\u0026nbsp;combined with SHAP performed the best for the quick and accurate identification of flood maximum inundation extent and their dominant drivers.\u003c/li\u003e\n \u003cli\u003eSHAP outperforms LIME by accounting for both direct contributions and interactions of driving factors.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"1 Introduction","content":"\u003cp\u003eCompound floods, driven by multiple interacting factors, are often more destructive than single-driver events. These events frequently cause substantial property damage and loss of life (Wahl et al. \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). The severe impacts of compound floods underscore the critical role of emergency response in protecting lives and property, particularly as flood patterns evolve (El baida et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). For example, in the United States, local Police and Fire Departments lead urban emergency flood responses, rapidly deploying resources to minimize harm (Yin et al., 2017). Similarly, the UK\u0026rsquo;s Environment Agency (EA) manages national flood emergency operations and deploying temporary defenses for numerous high-risk areas (Environment Agency, 2020). Such national efforts highlight that effective and timely emergency response is indispensable for urban flood resilience (Asif et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eRapid identification of inundation extent and dominant drivers provides the crucial basis for guiding effective emergency response in urban areas prone to compound floods. Predicting inundation extent aids timely evacuations (Yin et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and the deployment of temporary defenses (Qin et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e). Equally crucial is the identification of flood dominant drivers. In compound floods, various drivers (e.g., heavy rainfall, high river levels, storm surges, tides) interact or combine, leading to diverse flood types (Fang et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Because each flood type necessitates a tailored response strategy in practsice (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), pinpointing the dominant driver\u0026mdash;the primary factors determining the degree of flood inundation\u0026mdash;serves to guide effective and appropriate interventions. For example, when inundation results from insufficient drainage capacity due to intense local rainfall, localized measures\u0026mdash;such as opening stormwater inlets, excavating temporary channels, and constructing diversion pathways using temporary water barriers\u0026mdash;can effectively mitigate surface flooding. Conversely, when inundation is driven by high sea levels from storm surges or excessive river flow, which overwhelm drainage systems and cause backflow, localized measures alone are insufficient. In such scenarios, external emergency interventions, such as deploying mobile pumps and water storage tanks to extract and drain accumulated surface water (Qin et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e), as well as regulating downstream reservoirs or pumping stations to reduce river water levels, become essential. Thus, accurately identifying the dominant drivers of flood inundation is crucial for differentiating between locally manageable situations and those necessitating external support, thereby optimizing emergency resource allocation.\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 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eTypes of floods and corresponding mitigation measures\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\u003eCauses of flooding\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMitigation measures\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLiterature\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRainfall intensity exceeds the drainage capacity of the system\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eOpening stormwater inlets, excavating temporary channels, constructing diversion pathways with temporary water barriers, mobile pumps and mobile water storage tanks\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e(Qin et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e(Qin et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e(Shen et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUpstream runoff exceeding the river's capacity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRegulation of reservoirs or pumping stations to lower river water levels\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e(Sun et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2023\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e(Zhao et al. \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSea level rise and storm surges\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRegulation of tide gate, temporary water barriers\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e(Shen et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e(Han and Tahvildari \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\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\u003cp\u003eFlood inundation simulation typically relies on hydrodynamic models recognized for their accuracy and robust representation of underlying physical processes (Razavi et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). However, their significant computational demands often hinder practical application in real-time flood forecasting, despite various acceleration efforts (Rong et al. \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Sanders and Schubert \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). In contrast, machine learning (ML) models, including deep learning structures, can rapidly learn complex input-output relationships from training data. This capability enables accurate flood inundation predictions (Tian et al. \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). A spectrum of ML techniques has been investigated, from linear and tree-based algorithms to more sophisticated neural network architectures (Teng et al. \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Bentivoglio et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). For instance, Mangukiya et al.(2024) utilized the HEC-RAS framework to develop linear logistic regression, decision tree, and XGBoost models for simulating flood depth and inundated areas, reporting high accuracy and computational efficiency. Similarly, Kabir et al.(2020) employed a 1D Convolutional Neural Network for fluvial flood inundation modeling in Carlisle, UK, achieving prediction speeds 37 times faster than hydrodynamic model. These advancements highlight the potential of machine learning (ML) models as powerful tools for real-time flood inundation prediction.\u003c/p\u003e\u003cp\u003eWhile machine learning models excel at the predictive tasks, their common 'black-box' nature poses a significant challenge when attempting to use them for the explicit identification of dominant drivers in specific inundated areas (Rudin \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Materia et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Existing methods for driver analysis also have significant shortcomings. First, analytical techniques like feature contribution analyses (e.g., Partial Dependence Plots) are limited because they typically provide only global assessments of driver importance across an entire study area, failing to offer the location-specific insights crucial for targeted interventions (Zhou et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Rohmer et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Second, Physics-based methods, including hydrodynamic modeling combined with scenario analysis, are also commonly used to identify dominant flood drivers. Although these methods can effectively attribute flooding to different drivers, they are computationally intensive, limiting their applicability to long-term planning rather than real-time emergency management (Shen et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Xu et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Therefore, there is a lack of effective methods capable of both rapidly predicting flood inundation and concurrently providing location-specific identification of its dominant drivers.\u003c/p\u003e\u003cp\u003eThe emergence of explainable Artificial Intelligence (XAI) offers promising avenues to address the interpretability limitations inherent in 'black-box' models. By integrating XAI techniques with machine learning (ML) models, explainable machine learning models make it possible to concurrently generate predictions and quantify the contributions of drivers for specific instances (Liu et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Prominent methods such as SHAP (Lundberg and Lee \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) and LIME (Ribeiro et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) are capable of providing these explanations for individual predictions. For instance, Yang et al.(2020) utilized Shapley values to identify rainfall, snowmelt, and wetness as key drivers of peak flow in numerous Chinese catchments. Zahura and Goodall (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) employed LIME to show that local features like elevation and hourly rainfall significantly impacted flood depth estimations, with contributions varying by flood event type. Furthermore, Jiang et al.(2024) used SHAP to quantify the relative importance of interacting drivers in diverse flood magnitudes across thousands of global basins.\u003c/p\u003e\u003cp\u003eDespite the exploratory use of explainable machine learning models for flood inundation simulation and associated driver analysis in some case studies, a systematic evaluation is needed to determine how to best pair an XAI method with a high-performing ML model to ensure both predictive accuracy and robust interpretability.\u003c/p\u003e\u003cp\u003eThis paper systematically compares various explainable machine learning models for predicting maximum inundation extents and identifying their dominant drivers. We developed and evaluated a set of explainable machine learning models (EML). We constructed these models by pairing two representative XAI techniques, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), with three distinct types of ML models: a linear model (Logistic Regression), a tree-based ensemble (Random Forest), and a neural network (Multilayer Perceptron). The study area of Norfolk, Virginia, USA, was selected to train and test the models, using data from 26 historical compound flood events driven by heavy rainfall and high tides. This research aims to demonstrate the value of explainable machine learning methods in urban flood analysis, guiding informed decision-making and effective flood management in the face of growing compound flood risks, thereby advancing more targeted flood emergency management.\u003c/p\u003e"},{"header":"2 Study area and Data","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Study area\u003c/h2\u003e\u003cp\u003eNorfolk, Virginia (USA), is a coastal city in southeastern Virginia, situated at the confluence of the Elizabeth River and Chesapeake Bay (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). With an extensive 144 km shoreline, the city serves as a significant port. Norfolk's topography is exceptionally low-lying, with an average elevation of only 1.8 meters and a maximum of 12 meters. This makes it one of the U.S. coastal cities most susceptible to flooding, frequently experiencing inundation from storm surges and rising sea levels. Furthermore, the August-September period often brings increased tropical cyclone activity (hurricanes and tropical storms), resulting in heavy rainfall. The combination of these coastal and rainfall-driven factors leads to frequent compound flooding, posing a persistent challenge to the city (Zahura et al. \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Given Norfolk's vulnerability to such events and the availability of comprehensive street-scale meteorological, topographic, and water depth data ( Zahura and Goodall, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), this study focuses on an approximately 56.4 km\u0026sup2; urbanized area in the southwestern coastal part of the city.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Data preparation\u003c/h2\u003e\u003cdiv id=\"Sec5\" class=\"Section3\"\u003e\u003ch2\u003e2.2.1 Deriving maximum inundation extent for road segments\u003c/h2\u003e\u003cp\u003eIn the study area, flooding is primarily caused by rainfall and sea level rise. Flood events were identified through the selection of the top 16 rainfall events with the highest intensity and the top 10 tidal events with the highest sea levels, occurring from 2016 to 2018 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Figures A1 and A2).\u003c/p\u003e\u003cp\u003eThe water depth data utilized in this research were originally simulated using the TUFLOW model, a one-dimensional (1D) pipe network and two-dimensional (2D) overland hydrodynamic flood model. This model simulated inundation across the entire city of Norfolk for the 26 identified flood events (Zahura and Goodall \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The TUFLOW model provides water depth estimates at a 5-meter spatial resolution and a 1-hour temporal resolution. The model's calibration and validation have been previously established, ensuring the accuracy of its outputs (Shen et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Based on this existing research, hourly inundation depths were extracted for individual road segments. The road network was discretized into 16,868 segments, each 50 meters in length and 7.2 meters in width. Within each segment, flood conditioning factors and inundation depths were assumed to be homogeneous.\u003c/p\u003e\u003cp\u003eTo determine the maximum inundation extents, we converted the simulated hourly water depth sequences for each road segment into binary classifications (i.e., flooded or non-flooded) using a threshold-based approach. This process involved first extracting the hourly simulated water depth data for every segment across all 26 flood events. Subsequently, the maximum water depth experienced by each segment during each event was identified. A flooding threshold of 0.3 meters, corresponding to the approximate height of a typical passenger car's exhaust outlet, was then applied to classify segments as either flooded or non-flooded. This procedure resulted in the identification of 28,650 flooded road segment instances. The maximum inundation area for each event was subsequently calculated by multiplying the total number of flooded segments by the area of an individual segment.\u003c/p\u003e\u003cp\u003eTo ensure the representativeness of our training and testing datasets, we then employed an interval sampling method to select 13 events for the training set and the remaining 13 events for the testing set (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\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 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e26 rainfall and high tide flood events.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" 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=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFlood event date\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRainfall intensity averaged across all segments (mm/min)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMaximum tide level (m)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain or test\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eNumber of flooded segments\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e8/20/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e84.25\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.610\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1056\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e10/08/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e80.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.739\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e4130\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e9/21/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e37.66\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.734\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e2743\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e7/30/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e37.47\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.479\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e903\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e8/29/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e34.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.094\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1513\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e9/03/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e32.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.377\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e2384\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e7/15/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e32.67\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.441\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1684\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e7/31/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e31.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.475\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e2427\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e5/28/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e27.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.633\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1044\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e8/07/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e22.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.445\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e850\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e6/05/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e22.55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.624\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1569\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e1/02/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e21.28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.390\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e595\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e5/06/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e20.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.244\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1619\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e10/29/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e20.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.603\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e649\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e6/22/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e17.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.684\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1316\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e8/09/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e16.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.401\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e227\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e1/23/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e12.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.114\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1313\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e9/19/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e9.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.098\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e195\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e9/09/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5.31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.082\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e867\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e5/05/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.072\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e355\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e3/21/2018\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.021\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e386\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e9/29/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.012\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e159\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e2/09/2016\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e131\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e1/24/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.994\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e144\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e11/08/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.981\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e316\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003e1/23/2017\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.851\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e75\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=\"Sec6\" class=\"Section3\"\u003e\u003ch2\u003e2.2.2 Flood conditioning factors\u003c/h2\u003e\u003cp\u003eNine key flood conditioning factors were selected for this study, as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The selection was informed by a review of relevant literature, considering the specific characteristics of coastal urban environments and data availability within the study area. These factors are categorized into three groups: rainfall, tidal, and topographical.\u003c/p\u003e\u003cp\u003eFour rainfall-related factors were incorporated to characterize precipitation patterns:\u003c/p\u003e\u003cp\u003e1. Maximum 1-hour rainfall (Max_1HR): This represents the highest recorded rainfall intensity within a one-hour period for each road segment, reflecting the immediate impact of intense precipitation on surface inundation.\u003c/p\u003e\u003cp\u003e2. Cumulative rainfall in the first 2 hours (Max_1HR_2) and 72 hours (Max_1HR_72): These metrics quantify the total rainfall accumulated within 2 and 72 hours, respectively, relative to the timing of the peak one-hour rainfall for each flood event. Max_1HR_2 and Max_1HR_72 serve as proxies for the antecedent conditions, capturing the influence of recent and preceding rainfall on surface inundation.\u003c/p\u003e\u003cp\u003e3. Total hours of rainfall exceeding 0 mm (RH_Count): This factor provides information on the duration of a rainfall event. Longer rainfall durations are generally associated with an increased likelihood of surface flooding.\u003c/p\u003e\u003cp\u003eTwo tidal factors were utilized:\u003c/p\u003e\u003cp\u003e1. Maximum 1-hour tide level (Max_1TD): This denotes the peak tide level observed within a one-hour window for each road segment during a flood event, highlighting the direct influence of tidal surges on inundation.\u003c/p\u003e\u003cp\u003e2. Total hours tide level exceeded 0.8 m (TD_Count): In the study area, tidal flooding typically occurs when tide levels surpass 0.8 meters (Zahura and Goodall \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). TD_Count quantifies the duration for which this critical threshold is exceeded, with prolonged periods of high tides increasing the probability of surface inundation.\u003c/p\u003e\u003cp\u003eThree topographical features were included:\u003c/p\u003e\u003cp\u003e1. Elevation (EV): Previous research has consistently identified elevation as a crucial factor in flood susceptibility (Chapi et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Li et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), with lower-lying, flatter areas being inherently more prone to flooding than areas at higher elevations.\u003c/p\u003e\u003cp\u003e2. Distance to Water (DTW): This factor measures the shortest distance from a given road segment to the nearest surface water body. Smaller DTW values indicate closer proximity to water sources, which often correlates with wetter soils and a consequently higher likelihood of surface flood inundation (Lyu and Yin \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e3. Topographic Wetness Index (TWI): TWI is a steady-state wetness index that quantifies the potential for water to accumulate at any point in a catchment, thereby representing the propensity for water accumulation within the urban roadway network (Manfreda et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\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 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFactors used in flood inundation areas modeling for Norfolk\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=\"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\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\u003eNo\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eFeature/name (abbreviation)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u003cp\u003eUnit\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eData source\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eRainfall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMaximum one-hour rainfall (Max_1HR)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" morerows=\"3\" nameend=\"c6\" namest=\"c5\" rowspan=\"4\"\u003e\u003cp\u003eHourly rainfall observations were obtained from Hampton Roads Sanitation District (HRSD) observation sites\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCumulative rainfall for the previous 2hrs (Max_1HR_2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCumulative rainfall for the previous 72hrs (Max_1HR_72)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTotal hours of rainfall exceeding 0 mm (RH_Count)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003ehour\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eTidal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMaximum one-hour tide level (Max_1TD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003em\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" morerows=\"1\" nameend=\"c6\" namest=\"c5\" rowspan=\"2\"\u003e\u003cp\u003eHourly tide levels were obtained from NOAA's Sewells Point station\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTotal number of hours that the tide level exceeds 0.8m(TD_Count)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003ehour\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e\u003cp\u003eTopographic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eElevation (EV)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003em\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" morerows=\"2\" nameend=\"c6\" namest=\"c5\" rowspan=\"3\"\u003e\u003cp\u003eA 5 m Digital Elevation Model (DEM) was obtained from the U.S. Geological Survey (USGS)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDepth to water index (DTW)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003ecm\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTopographic wetness index (TWI)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\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\u003c/div\u003e"},{"header":"3 Methods","content":"\u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the research workflow. It begins with the data preparation phase (detailed in the previous section). Subsequently, the explainable machine learning stage employs the training data to develop flood inundation predictive models, specifically Linear Regression (LR), Random Forest (RF), and Multilayer Perceptron (MLP). Explainable AI techniques were then employed to interpret the predictions generated by these models and to determine the dominant flood drivers, followed by an evaluation of the accuracy of the flood inundation predictions and the dominant driver identifications.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Maximum inundation extent prediction\u003c/h2\u003e\u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\u003ch2\u003e3.1.1 Machine learning models\u003c/h2\u003e\u003cp\u003eIn this study, we selected representative models from each type to model flood maximum inundation areas: LR (linear), RF (tree-based), and MLP (network-based).\u003c/p\u003e\u003cp\u003eLogistic regression (LR), introduced by Cox (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e1958\u003c/span\u003e), is a linear statistical model that establishes a relationship between a set of predictor variables (i.e., conditioning factors) and a binary dependent variable. Pradhan (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) first applied LR to flood inundation area modeling. In this context, the LR model is trained on historical data to discern the relationship between various geospatial and event-specific features and binary flood labels (e.g., flooded/non-flooded). The model employs a sigmoid function to transform a linear combination of input features into a probability estimate (ranging from 0 to 1) of flooding for each segment.\u003c/p\u003e\u003cp\u003eRandom Forest (RF), proposed by Breiman (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2001\u003c/span\u003e), is an ensemble tree-based algorithm widely adopted for flood inundation area modeling since its initial application in this field by Trigila et al.(2015). RF constructs multiple decision trees during training and outputs the mode of the classes (for classification) or mean prediction (for regression) of the individual trees, thereby enhancing predictive accuracy and stability. In flood inundation modeling, RF offers several advantages, including robustness to multicollinearity among predictor variables, effective handling of missing data, and a significant reduction in overfitting compared to individual decision trees (Choubin et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These characteristics make it particularly well-suited for complex flood prediction tasks.\u003c/p\u003e\u003cp\u003eAdditionally, neural network algorithms have been adopted in flood modelling due to their strong predictive capabilities (Zhao et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Various neural network models, including MLP, CNN, RNN and other more advanced deep learning models, have been widely used in this field (Tien Bui et al. \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Among these, MLP is the fundamental algorithm, consisting of multiple layers of computational units (i.e., nodes or neurons) connected by weights. In the MLP approach, input-output relationships are established through multiple layers of nonlinear transformations. The strength of neural network models lies in their ability to address complex nonlinear problems.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e3.1.2 Model training\u003c/h2\u003e\u003cp\u003eThe training process for the models involves hyperparameter optimization and implementation. Each model has several key hyperparameters that greatly influence performance, making tuning a crucial step. Hyperparameter tuning was conducted via 5-fold cross-validation using \"GridSearchCV\" for automatic optimization. Given the focus on accurate flood inundation detection, recall was used as the criterion for optimization. Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e provides the optimal hyperparameters for all models.\u003c/p\u003e\u003cp\u003eThe construction and training of LR, RF, and MLP models were implemented using the LogisticRegression, RandomForestRegressor, and MLPClassifier modules of the scikit-learn library in Python.\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\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDescription of models and their hyperparameters.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHyper-parameters\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRange\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eOptimized value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLinear\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003epenalty\u003c/p\u003e\u003cp\u003ec\u003c/p\u003e\u003cp\u003esolver\u003c/p\u003e\u003cp\u003emax_iter\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e[\u0026lsquo;L1\u0026rsquo;, \u0026lsquo;L2\u0026rsquo;]\u003c/p\u003e\u003cp\u003e[0.1, 1, 10]\u003c/p\u003e\u003cp\u003e[\u0026lsquo;liblinear\u0026rsquo;, \u0026lsquo;saga\u0026rsquo;]\u003c/p\u003e\u003cp\u003e[100, 200, 500]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eL1\u003c/p\u003e\u003cp\u003e10\u003c/p\u003e\u003cp\u003eliblinear\u003c/p\u003e\u003cp\u003e100\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eTree\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003en_estimators\u003c/p\u003e\u003cp\u003emax_depth\u003c/p\u003e\u003cp\u003emin_samples_split\u003c/p\u003e\u003cp\u003emin_samples_leaf\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e[50, 100, 200]\u003c/p\u003e\u003cp\u003e[5, 10, 20, None]\u003c/p\u003e\u003cp\u003e[2, 5, 10]\u003c/p\u003e\u003cp\u003e[1, 2, 4]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e150\u003c/p\u003e\u003cp\u003eNone\u003c/p\u003e\u003cp\u003e5\u003c/p\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eccp_alpha\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e[0.0002, 0.00025, 0.0003]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.00025\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNeural network\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMLP\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ehidden_layer_sizes\u003c/p\u003e\u003cp\u003eactivation\u003c/p\u003e\u003cp\u003esolver\u003c/p\u003e\u003cp\u003elearning_rate\u003c/p\u003e\u003cp\u003emax_iter\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e[(10,), (50,), (100,)]\u003c/p\u003e\u003cp\u003e[\u0026lsquo;relu\u0026rsquo;, \u0026lsquo;tanh\u0026rsquo;]\u003c/p\u003e\u003cp\u003e[\u0026lsquo;sgd\u0026rsquo;, \u0026lsquo;adam\u0026rsquo;]\u003c/p\u003e\u003cp\u003e[\u0026lsquo;constant\u0026rsquo;, \u0026lsquo;adaptive\u0026rsquo;]\u003c/p\u003e\u003cp\u003e[100, 200, 500]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e(100,)\u003c/p\u003e\u003cp\u003etanh\u003c/p\u003e\u003cp\u003eadam\u003c/p\u003e\u003cp\u003econstant\u003c/p\u003e\u003cp\u003e500\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\u003e3.1.3 Evaluation metrics\u003c/h2\u003e\u003cp\u003eThe performance of the machine learning models was evaluated using several standard metrics, including the Receiver Operating Characteristic (ROC) curve, the Area Under the Curve (AUC), accuracy, precision, recall, and the F1 score, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. These metrics are derived from the classification outcomes for each road segment, which can be categorized into four types based on a confusion matrix. True Positive (TP): Represents the number of road segments correctly predicted as flooded. False Positive (FP): Represents the number of non-flooded road segments incorrectly predicted as flooded. True Negative (TN): Represents the number of non-flooded road segments correctly predicted as non-flooded. False Negative (FN): Represents the number of flooded road segments incorrectly predicted as non-flooded.\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 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModel Performance Evaluation Metrics\u003c/p\u003e\u003c/div\u003e\u003c/caption\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\u003eEvaluation criterion\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDescription \u0026amp; formula\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTure positive (TP)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNumber of flood samples correctly classified.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFalse Positive (FP)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNumber of flood samples not correctly classified.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTrue Negative (TN)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNumber of non-flood samples correctly classified.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFalse Negative (FN)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNumber of non-flood samples not correctly classified.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAccuracy (Acc)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Accuracy=\\:\\frac{TP+TN}{TP+TN+FP+FN}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall (R)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Recall=\\:\\frac{TP}{TP+FN}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrecision (P)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Precision=\\:\\frac{TP}{TP+FP}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1 score (F)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:F1\\:score=2\\times\\:\\frac{P\\times\\:R}{P+R}\\)\u003c/span\u003e\u003c/span\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\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Dominant flood driver identification\u003c/h2\u003e\u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\u003ch2\u003e3.2.1 Explainable AI (XAI) Techniques\u003c/h2\u003e\u003cp\u003eTo identify the dominant drivers of flood inundation, this study employed two widely recognized, model-agnostic explainable AI (XAI) techniques: Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eLIME, developed by Ribeiro et al.(2016), is designed to explain the predictions of any machine learning model for individual instances. It operates by approximating the behavior of a complex model locally with a simpler, more interpretable model (e.g., a linear model). The core assumption is that while the global behavior of a complex model may be intricate, its decision-making process for a specific instance can be faithfully represented by a simpler model learned in the vicinity of that instance. Operationally, LIME generates a new dataset of perturbed samples around the instance of interest, obtains predictions for these samples from the original complex model, and then trains an interpretable model on this new dataset weighted by proximity to the original instance. The parameters of this local interpretable model (such as the weights in a linear regression) then serve as explanations for the individual prediction, indicating the contribution of each feature to that specific outcome. This approach enhances the interpretability of complex models such as random forests and neural networks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSHAP, introduced by Lundberg and Lee (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), is another prominent technique for explaining model predictions. It leverages Shapley values, a concept from cooperative game theory, to attribute the prediction outcome to each feature. A feature's Shapley value represents its average marginal contribution to the prediction across all possible combinations of features, thereby capturing both its individual impact and its interactions with other features. For a given prediction, the sum of the SHAP values for all features, plus a baseline value (typically the average prediction over the training dataset), equals the model's actual output for that instance. SHAP values thus provide a consistent and locally accurate measure of feature importance, ensuring that the explanation is fair and reliable.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e3.2.2 Identifying dominant flood drivers using XAI techniques\u003c/h2\u003e\u003cp\u003eThe process for identifying dominant flood drivers using the selected XAI techniques involved the following steps:\u003c/p\u003e\u003cp\u003e1. Feature contribution calculation: Subsequent to the training of the machine learning models, both LIME and SHAP were applied to interpret the model predictions. This involved calculating the specific contribution of each input feature to the predicted flood status for every individual road segment within each flood event.\u003c/p\u003e\u003cp\u003e2. Extraction of driver-specific contributions: From the comprehensive set of feature contributions, values corresponding to the four designated rainfall features and the two designated tidal features were specifically extracted for further analysis.\u003c/p\u003e\u003cp\u003e3. Dominant driver classification: For each road segment predicted as inundated, the feature (from the aforementioned rainfall and tidal categories) with the highest positive contribution to the flood prediction was identified. Based on whether this most influential feature was rainfall-related or tidal-related, the inundated road segment was classified as either rainfall-dominated or tidal-dominated.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"4 Results","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Predictive performance of ML models\u003c/h2\u003e\u003cp\u003eThe predictive performance of machine learning models for maximum flood inundation was assessed using Accuracy, Precision, Recall, F1 score, and ROC-AUC, with results summarized in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. RF model consistently outperformed MLP and LR.\u003c/p\u003e\u003cp\u003eOn the testing set, RF achieved the highest Accuracy (0.79, compared to MLP's 0.75 and LR's 0.74) and Recall (0.81, vs. MLP 0.78, LR 0.67), demonstrating its effectiveness in minimizing the omission of true flood events. For Precision, RF and LR (both 0.78) surpassed MLP (0.74), indicating reliable positive predictions when a segment was classified as flooded. RF also led with the top F1 score of 0.79 (MLP 0.76, LR 0.72). Collectively, these results underscore RF's superior robustness and reliability in predicting maximum flood inundation, establishing it as the most effective model among those evaluated.\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 6\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePredictive performance evaluation of each ML model.\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=\"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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u003cp\u003eTraining\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c8\" namest=\"c6\"\u003e\u003cp\u003eTesting\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMLP\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eMLP\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.67\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.81\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1 score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.76\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\u003eROC curve analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) further highlighted the RF model's superior performance. On the test dataset, RF achieved the highest AUC with a value of 0.870, outperforming both MLP (AUC 0.829) and LR (AUC 0.809). Considering its consistently strong performance across all evaluation metrics, the RF model was therefore selected for the subsequent identification of dominant flood drivers.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Predicted maximum flood inundation maps\u003c/h2\u003e\u003cp\u003eGiven the extensive number of flood events in the test set, three representative events were selected for a detailed spatial analysis of RF model's predictions. These events were chosen based on differing rainfall intensity and tidal levels: (1) A high rainfall intensity, low tidal level event (July 31, 2016), (2) A low rainfall intensity, high tidal level event (January 24, 2017), (3) A high rainfall intensity, high tidal level event (September 3, 2016). Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates the spatial distribution of the predicted maximum flood inundation for these events, with prediction outcomes categorized by True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN).\u003c/p\u003e\u003cp\u003eFor the high rainfall, low tidal level event (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea-b), the RF model correctly identified 1,874 out of 2,427 actual inundated road segments, achieving a recall of approximately 77%. The inundated segments were widely distributed, with notable concentrations in the southwestern part of the study area, particularly in downtown Norfolk.\u003c/p\u003e\u003cp\u003eIn the low rainfall, high tidal level event (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec-d), the model demonstrated high performance, correctly predicting 140 out of 144 actual inundated segments. This resulted in a recall of approximately 97%. Unlike the widespread inundation in the high-rainfall event, the flooded segments in this scenario were more sparsely distributed and primarily located near the central coastal area.\u003c/p\u003e\u003cp\u003eFor the high rainfall, high tidal level event (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee-f), the model correctly identified 1,806 out of 2,384 actual inundated segments, corresponding to a recall of approximately 76%. The correctly predicted segments were extensively distributed across both coastal and inland areas. This widespread inundation likely resulted from the combined influence of significant rainfall and high tidal levels, which jointly increased the probability of flooding.\u003c/p\u003e\u003cp\u003eOverall, the visual and quantitative analysis of these representative events (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) demonstrates the RF model's robust capability in predicting the spatial extent of flood inundation under varied environmental conditions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Dominant flood driver analysis: Comparing SHAP and LIME\u003c/h2\u003e\u003cp\u003eThe SHAP and LIME techniques were utilized to analyze the contributions of various rainfall and tidal features to the flood predictions made by the RF model, specifically for those road segments correctly identified as inundated. The dominant driver (rainfall or tidal) for each such segment was determined by identifying whether the feature exhibiting the highest positive contribution was rainfall-related or tidal-related.\u003c/p\u003e\u003cp\u003eGiven the lack of direct ground-truth observations detailing the actual dominant flood driver for each road segment during each historical event, this study assesses the spatial reasonableness of the derived driver distributions to compare the effectiveness of the different methods. The evaluation is based on the premise that a plausible spatial pattern would generally conform to expected hydrological behaviors in coastal cities. Specifically, it is assumed that rainfall-dominated inundation would likely be more dispersed throughout the study area, whereas tidal-dominated inundation would primarily be concentrated along the coastline and tidally influenced water bodies. The consistency of the XAI-derived driver patterns with these expected distributions serves as an indicator of their rationality.\u003c/p\u003e\u003cp\u003eThe comparative results for three representative flood events are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. During the January 24, 2017 event (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb, e), characterized by high tidal levels and low rainfall, SHAP and LIME again demonstrated close agreement. Both methods predominantly classified the correctly predicted flooded segments as tide-dominated (SHAP: 136/140, ~\u0026thinsp;97.1%; LIME: 138/140, ~\u0026thinsp;98.6%). For both XAI tools, TD_Count was overwhelmingly the principal feature identified for these tide-dominated instances. As anticipated, these segments were primarily concentrated along the coast, consistent with tidal influence patterns.\u003c/p\u003e\u003cp\u003eSignificant difference between SHAP and LIME occurred for the September 3, 2016 compound event (high rainfall and high tidal levels; Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec,f), which involved 1,806 correctly predicted flooded segments.\u003c/p\u003e\u003cp\u003eSHAP classified 1,168 segments (approx. 64.7%) as rainfall-dominated and 638 segments (approx. 35.3%) as tide-dominated. Among SHAP's rainfall-dominated segments, Max_1HR_72 was the leading contributor (approx. 51%), followed by RH_Count (approx. 8%) and Max_1HR_2 (approx. 5%). For SHAP's tide-dominated segments, TD_Count was the dominant feature in approximately 96% of cases. LIME classified approximately 1,427 segments (79% of 1,806) as rainfall-dominated and approximately 379 segments (21% of 1,806) as tide-dominated. LIME attributed about 68% of its rainfall-dominated segments to Max_1HR_72, with TD_Count and RH_Count also noted as significant contributors. For LIME's tide-dominated segments, TD_Count reportedly influenced 66% and Max_1TD influenced 34%.\u003c/p\u003e\u003cp\u003eSpatially, SHAP\u0026rsquo;s results for the compound event (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec) depicted tide-dominated segments clustered near the coast and rainfall-dominated segments dispersed throughout the study area, which is consistent with expected hydrological patterns. In contrast, LIME\u0026rsquo;s results (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ef) showed a less intuitive distribution, placing many tide-dominated segments inland (particularly in the northern region) and concentrating rainfall-dominated segments along the coast, thereby deviating from empirical patterns and the physical understanding of these drivers.\u003c/p\u003e\u003cp\u003eOverall, the results obtained from SHAP demonstrate a closer alignment with expected flood patterns in coastal cities.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"5 Discussions","content":"\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Effectiveness of explainable ML models\u003c/h2\u003e\u003cp\u003eThis study evaluated three ML models\u0026mdash;LR, RF, and MLP\u0026mdash;along with two XAI techniques, LIME and SHAP. The evaluation centered on two key aspects: the accuracy of flood inundation area prediction and the spatial reasonableness of dominant flood driver identification. Among the ML model-XAI combinations assessed, the RF model paired with SHAP yielded the most robust performance.\u003c/p\u003e\u003cp\u003eRegarding predictive capability for flood inundation, the tree-structured RF model surpassed the other models, achieving an accuracy of 0.79 in discriminating between flooded and non-flooded road segments. This superior performance of RF is consistent with findings from similar comparative studies in flood modeling (Rezvani et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Conversely, the LR model exhibited the weakest predictive power. This outcome also aligns with its frequent use as a benchmark in existing literature, where it often serves as a baseline for comparison against more sophisticated models (Hong et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Al-Juaidi \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eFrom the perspective of identifying dominant flood drivers, this study employed XAI techniques to identify dominant drivers at specific inundation locations. This approach contrasts with many previous studies that typically offer global feature importance rankings applicable to the entire study area. XAI techniques delivers more granular, localized insights into flood causality (Avand et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Mahdizadeh Gharakhanlou and Perez \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), which can be more actionable for targeted flood mitigation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Rationale for RF-SHAP's effectiveness\u003c/h2\u003e\u003cp\u003eThe study's results demonstrate that RF combined with SHAP achieved superior performance in both predicting flood inundation areas (Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e) and identifying dominant flood drivers (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). The reasons for this are as follows.\u003c/p\u003e\u003cp\u003eRF's effectiveness in flood extent prediction lies in its ensemble structure, which robustly captures the complex non-linear relationships inherent in flood phenomena, often with less extensive data and tuning requirements than models like MLP (Tyralis et al. \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eFor dominant flood driver identification, SHAP demonstrated advantages over LIME, particularly in accurately representing driver influence during complex compound flood events (e.g., Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec,f). The primary reason is SHAP's methodology, which inherently accounts for feature interactions by evaluating how the contribution of one feature is affected by the presence or absence of others (Lundberg and Lee \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). This is crucial because flood drivers (like rainfall and tidal influences) are often interdependent. LIME, by typically relying on local linear approximations and assuming feature independence, is less adept at capturing these critical interaction effects (Ribeiro et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Given that the interplay between different factors significantly influences flood occurrence and characteristics, SHAP\u0026rsquo;s ability to incorporate these interactions into its attributions provides a more reliable and mechanistically sound identification of dominant flood drivers compared to LIME in such complex environmental systems.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003e5.3 Leveraging explainable ML model for enhanced flood emergency response\u003c/h2\u003e\u003cp\u003eExplainable machine learning models like RF-SHAP, validated in this study, offer significant practical value for flood emergency response. By identifying location-specific dominant flood drivers (e.g., tidal vs. rainfall), the method helps authorities develop more targeted interventions and optimize resource allocation (Cools et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). It also overcomes the \"black box\" problem of traditional AI by explaining the reasoning behind its predictions, which builds trust and leads to more confident, science-based decisions by managers (Adadi and Berrada \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Crucially, the model is highly efficient, providing near real-time prediction and attribution (approx. 33 seconds per event), which is essential for managing rapidly changing flood conditions.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section2\"\u003e\u003ch2\u003e5.4 Limitations and future work\u003c/h2\u003e\u003cp\u003eDespite the demonstrated effectiveness of the RF-SHAP approach for flood analysis in Norfolk, Virginia, this study has limitations that suggest avenues for future research. A primary limitation is that the methodology's validation is currently geographically specific. Future work should therefore evaluate its performance and adaptability in diverse regions with varying hydrogeological and urban characteristics to ascertain broader generalizability. Furthermore, this study focused on maximum flood inundation, not the dynamic spatiotemporal evolution of flooding. Future research could aim to develop models that dynamically predict flood progression over time and space, concurrently identifying how dominant drivers evolve throughout a flood event. This would offer a more comprehensive understanding, particularly valuable for real-time risk assessment and response.\u003c/p\u003e\u003c/div\u003e"},{"header":"6 Conclusions","content":"\u003cp\u003eFor effective emergency response in urban areas vulnerable to compound flooding, the rapid identification of inundation areas and their dominant drivers is essential. This study addressed this dual challenge by systematically evaluating explainable machine learning (EML) frameworks that pair machine learning (ML) models with explainable AI (XAI) techniques.\u003c/p\u003e\u003cp\u003eOur principal findings show that the combination of Random Forest (RF) and SHAP is the most effective approach for both rapid flood prediction and reliable driver identification. The superiority of this pairing stems from SHAP's ability to accurately quantify driver influence, even in complex compound events, by accounting for the critical interactions between flood factors\u0026mdash;a key limitation of other methods like LIME, which largely assume feature independence.\u003c/p\u003e\u003cp\u003eThe integration of the RF-SHAP model into flood management offers a powerful tool to substantially enhance emergency response. By providing rapid, spatially-specific inundation predictions and interpretable insights into dominant drivers, this approach enables more targeted strategies, optimized resource allocation, and confident decision support. These advancements are crucial for building resilience in coastal cities facing growing compound flood risks.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Natural Science Foundation of China (Grant No. 52309005), the Fundamental Research Funds for the Central Universities (Grant No. DUT24ZD409), the National Natural Science Foundation of China (Grant No. U2240204), and the National Key R\u0026amp;D Program of China (Grant No. 2024YFC3213000). We would also like to thank Dr. Faria Zahura and her team for providing the foundational data that greatly facilitated our research.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Natural Science Foundation of China (Grant No. 52309005), the Fundamental Research Funds for the Central Universities (Grant No. DUT24ZD409), the National Natural Science Foundation of China (Grant No. U2240204), and the National Key R\u0026amp;D Program of China (Grant No. 2024YFC3213000).\u003c/p\u003e\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Jiqiang Xie: Writing \u0026ndash; original draft, Methodology, Data curation, Conceptualization. Heng Lyu: Writing \u0026ndash; review \u0026amp; editing, Funding acquisition, Conceptualization. Bing Yu: Review \u0026amp; editing, Conceptualization. Shengnan Fu: Writing \u0026ndash; review \u0026amp; editing. Chen Yang: Methodology, Conceptualization. Chi Zhang: Supervision.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAdadi A, Berrada M (2018) Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access 6:52138\u0026ndash;52160. https://doi.org/10.1109/ACCESS.2018.2870052\u003c/li\u003e\n\u003cli\u003eAl-Juaidi AEM (2023) The interaction of topographic slope with various geo-environmental flood-causing factors on flood prediction and susceptibility mapping. Environ Sci Pollut Res 30:59327\u0026ndash;59348. https://doi.org/10.1007/s11356-023-26616-y\u003c/li\u003e\n\u003cli\u003eAsif M, Kuglitsch MM, Pelivan I, Albano R (2025) Review and Intercomparison of Machine Learning Applications for Short-term Flood Forecasting. Water Resour Manag 39:1971\u0026ndash;1991. https://doi.org/10.1007/s11269-025-04093-x\u003c/li\u003e\n\u003cli\u003eAvand M, Moradi H, lasboyee MR (2021) Spatial modeling of flood probability using geo-environmental variables and machine learning models, case study: Tajan watershed, Iran. Adv Space Res 67:3169\u0026ndash;3186. https://doi.org/10.1016/j.asr.2021.02.011\u003c/li\u003e\n\u003cli\u003eBentivoglio R, Isufi E, Jonkman SN, Taormina R (2022) Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrol Earth Syst Sci 26:4345\u0026ndash;4378. https://doi.org/10.5194/hess-26-4345-2022\u003c/li\u003e\n\u003cli\u003eBreiman L (2001) Random Forests. Mach Learn 45:5\u0026ndash;32. https://doi.org/10.1023/A:1010933404324\u003c/li\u003e\n\u003cli\u003eChapi K, Singh VP, Shirzadi A, et al (2017) A novel hybrid artificial intelligence approach for flood susceptibility assessment. Environ Model Softw 95:229\u0026ndash;245. https://doi.org/10.1016/j.envsoft.2017.06.012\u003c/li\u003e\n\u003cli\u003eChoubin B, Moradi E, Golshan M, et al (2019) An ensemble prediction of flood susceptibility using multivariate discriminant analysis, classification and regression trees, and support vector machines. Sci Total Environ 651:2087\u0026ndash;2096. https://doi.org/10.1016/j.scitotenv.2018.10.064\u003c/li\u003e\n\u003cli\u003eCools J, Innocenti D, O\u0026rsquo;Brien S (2016) Lessons from flood early warning systems. Environ Sci Policy 58:117\u0026ndash;122. https://doi.org/10.1016/j.envsci.2016.01.006\u003c/li\u003e\n\u003cli\u003eCox DR (1958) The Regression Analysis of Binary Sequences. J R Stat Soc Ser B Methodol 20:215\u0026ndash;232. https://doi.org/10.1111/j.2517-6161.1958.tb00292.x\u003c/li\u003e\n\u003cli\u003eEl baida M, Boushaba F, Chourak M, Hosni M (2024) Real-Time Urban Flood Depth Mapping: Convolutional Neural Networks for Pluvial and Fluvial Flood Emulation. Water Resour Manag 38:4763\u0026ndash;4782. https://doi.org/10.1007/s11269-024-03886-w\u003c/li\u003e\n\u003cli\u003eFang J, Wahl T, Fang J, et al (2021) Compound flood potential from storm surge and heavy precipitation in coastal China: dependence, drivers, and impacts. Hydrol Earth Syst Sci 25:4403\u0026ndash;4416. https://doi.org/10.5194/hess-25-4403-2021\u003c/li\u003e\n\u003cli\u003eHan S, Tahvildari N (2024) Compound Flooding Hazards Due To Storm Surge and Pluvial Flow in a Low-Gradient Coastal Region. Water Resour Res 60:e2023WR037014. https://doi.org/10.1029/2023WR037014\u003c/li\u003e\n\u003cli\u003eHendry A, Haigh ID, Nicholls RJ, et al (2019) Assessing the characteristics and drivers of compound flooding events around the UK coast. Hydrol Earth Syst Sci 23:3117\u0026ndash;3139. https://doi.org/10.5194/hess-23-3117-2019\u003c/li\u003e\n\u003cli\u003eHong H, Tsangaratos P, Ilia I, et al (2018) Application of fuzzy weight of evidence and data mining techniques in construction of flood susceptibility map of Poyang County, China. Sci Total Environ 625:575\u0026ndash;588. https://doi.org/10.1016/j.scitotenv.2017.12.256\u003c/li\u003e\n\u003cli\u003eJiang S, Tarasova L, Yu G, Zscheischler J (2024) Compounding effects in flood drivers challenge estimates of extreme river floods. Sci Adv 10:eadl4005. https://doi.org/10.1126/sciadv.adl4005\u003c/li\u003e\n\u003cli\u003eKabir S, Patidar S, Xia X, et al (2020) A deep convolutional neural network model for rapid prediction of fluvial flood inundation. J Hydrol 590:125481. https://doi.org/10.1016/j.jhydrol.2020.125481\u003c/li\u003e\n\u003cli\u003eLi Y, Osei FB, Hu T, Stein A (2023) Urban flood susceptibility mapping based on social media data in Chengdu city, China. Sustain Cities Soc 88:104307. https://doi.org/10.1016/j.scs.2022.104307\u003c/li\u003e\n\u003cli\u003eLiu Y, Wang H, Guan X, et al (2025) Urban Flood Depth Prediction and Visualization Based on the XGBoost-SHAP Model. Water Resour Manag 39:1353\u0026ndash;1375. https://doi.org/10.1007/s11269-024-04020-6\u003c/li\u003e\n\u003cli\u003eLundberg SM, Lee S-I (2017) A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Curran Associates Inc., Red Hook, NY, USA, pp 4768\u0026ndash;4777\u003c/li\u003e\n\u003cli\u003eLyu H-M, Yin Z-Y (2023) Flood susceptibility prediction using tree-based machine learning models in the GBA. Sustain Cities Soc 97:104744. https://doi.org/10.1016/j.scs.2023.104744\u003c/li\u003e\n\u003cli\u003eMahdizadeh Gharakhanlou N, Perez L (2023) Flood susceptible prediction through the use of geospatial variables and machine learning methods. J Hydrol 617:129121. https://doi.org/10.1016/j.jhydrol.2023.129121\u003c/li\u003e\n\u003cli\u003eManfreda S, Nardi F, Samela C, et al (2014) Investigation on the use of geomorphic approaches for the delineation of flood prone areas. J Hydrol 517:863\u0026ndash;876. https://doi.org/10.1016/j.jhydrol.2014.06.009\u003c/li\u003e\n\u003cli\u003eMangukiya NK, Kushwaha S, Sharma A (2024) A novel multi-model ensemble framework for fluvial flood inundation mapping. Environ Model Softw 180:106163. https://doi.org/10.1016/j.envsoft.2024.106163\u003c/li\u003e\n\u003cli\u003eMateria S, Garc\u0026iacute;a LP, van Straaten C, et al (2024) Artificial intelligence for climate prediction of extremes: State of the art, challenges, and future perspectives. WIREs Clim Change 15:e914. https://doi.org/10.1002/wcc.914\u003c/li\u003e\n\u003cli\u003ePradhan B (2009) Flood susceptible mapping and risk area delineation using logistic regression, GIS and remote sensing. J Spat Hydrol 9:1\u0026ndash;18\u003c/li\u003e\n\u003cli\u003eQin H, Liang Q, Chen H, De Silva V (2024a) A Coupled Human and Natural Systems (CHANS) framework integrated with reinforcement learning for urban flood mitigation. J Hydrol 643:131918. https://doi.org/10.1016/j.jhydrol.2024.131918\u003c/li\u003e\n\u003cli\u003eQin H, Liang Q, Chen H, De Silva V (2024b) A two-way coupled CHANS model for flood emergency management, with a focus on temporary flood defences. Environ Model Softw 181:106166. https://doi.org/10.1016/j.envsoft.2024.106166\u003c/li\u003e\n\u003cli\u003eRazavi S, Tolson BA, Burn DH (2012) Review of surrogate modeling in water resources. Water Resour Res 48:. https://doi.org/10.1029/2011WR011527\u003c/li\u003e\n\u003cli\u003eRezvani SM, Gon\u0026ccedil;alves A, Silva MJF, de Almeida NM (2024) Smart hotspot detection using geospatial artificial intelligence: A machine learning approach to reduce flood risk. Sustain Cities Soc 115:105873. https://doi.org/10.1016/j.scs.2024.105873\u003c/li\u003e\n\u003cli\u003eRibeiro MT, Singh S, Guestrin C (2016) \u0026ldquo;Why Should I Trust You?\u0026rdquo;: Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, New York, NY, USA, pp 1135\u0026ndash;1144\u003c/li\u003e\n\u003cli\u003eRohmer J, Idier D, Paris F, et al (2018) Casting light on forcing and breaching scenarios that lead to marine inundation: Combining numerical simulations with a random-forest classification approach. Environ Model Softw 104:64\u0026ndash;80. https://doi.org/10.1016/j.envsoft.2018.03.003\u003c/li\u003e\n\u003cli\u003eRong Y, Bates P, Neal J (2024) GPU‐Accelerated Urban Flood Modeling Using a Nonuniform Structured Grid and a Super Grid Scale River Channel. Water Resour Res e2023WR036128: https://doi.org/10.1029/2023WR036128\u003c/li\u003e\n\u003cli\u003eRudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell 1:206\u0026ndash;215. https://doi.org/10.1038/s42256-019-0048-x\u003c/li\u003e\n\u003cli\u003eSanders BF, Schubert JE (2019) PRIMo: Parallel raster inundation model. Adv Water Resour 126:79\u0026ndash;95. https://doi.org/10.1016/j.advwatres.2019.02.007\u003c/li\u003e\n\u003cli\u003eShen Y, Morsy MM, Huxley C, et al (2019) Flood risk assessment and increased resilience for coastal urban watersheds under the combined impact of storm tide and heavy rainfall. J Hydrol 579:124159. https://doi.org/10.1016/j.jhydrol.2019.124159\u003c/li\u003e\n\u003cli\u003eSun J, Chen W, Hu B, et al (2023) Roles of reservoirs in regulating basin flood and droughts risks under climate change: Historical assessment and future projection. J Hydrol Reg Stud 48:101453. https://doi.org/10.1016/j.ejrh.2023.101453\u003c/li\u003e\n\u003cli\u003eTeng J, Jakeman AJ, Vaze J, et al (2017) Flood inundation modelling: A review of methods, recent advances and uncertainty analysis. Environ Model Softw 90:201\u0026ndash;216. https://doi.org/10.1016/j.envsoft.2017.01.006\u003c/li\u003e\n\u003cli\u003eTian Y, Cui R, Fu W, et al (2025) Dual-modality Encoder-decoder Framework for Urban Real-time Rainfall-runoff Prediction. Water Resour Manag 39:4509\u0026ndash;4538. https://doi.org/10.1007/s11269-025-04166-x\u003c/li\u003e\n\u003cli\u003eTien Bui D, Hoang N-D, Mart\u0026iacute;nez-\u0026Aacute;lvarez F, et al (2020) A novel deep learning neural network approach for predicting flash flood susceptibility: A case study at a high frequency tropical storm area. Sci Total Environ 701:134413. https://doi.org/10.1016/j.scitotenv.2019.134413\u003c/li\u003e\n\u003cli\u003eTrigila A, Iadanza C, Esposito C, Scarascia-Mugnozza G (2015) Comparison of Logistic Regression and Random Forests techniques for shallow landslide susceptibility assessment in Giampilieri (NE Sicily, Italy). Geomorphology 249:119\u0026ndash;136. https://doi.org/10.1016/j.geomorph.2015.06.001\u003c/li\u003e\n\u003cli\u003eTyralis H, Papacharalampous G, Langousis A (2019) A Brief Review of Random Forests for Water Scientists and Practitioners and Their Recent History in Water Resources. Water 11:910. https://doi.org/10.3390/w11050910\u003c/li\u003e\n\u003cli\u003eWahl T, Jain S, Bender J, et al (2015) Increasing risk of compound flooding from storm surge and rainfall for major US cities. Nat Clim Change 5:1093\u0026ndash;1097. https://doi.org/10.1038/nclimate2736\u003c/li\u003e\n\u003cli\u003eXu H, Ragno E, Jonkman SN, et al (2024) Combining statistical and hydrodynamic models to assess compound flood hazards from rainfall and storm surge: a case study of Shanghai. Hydrol Earth Syst Sci 28:3919\u0026ndash;3930. https://doi.org/10.5194/hess-28-3919-2024\u003c/li\u003e\n\u003cli\u003eYang W, Yang H, Yang D (2020) Classifying floods by quantifying driver contributions in the Eastern Monsoon Region of China. J Hydrol 585:124767. https://doi.org/10.1016/j.jhydrol.2020.124767\u003c/li\u003e\n\u003cli\u003eYin J, Yang Y, Yu D, et al (2024) Strategic storm flood evacuation planning for large coastal cities enables more effective transfer of elderly populations. Nat Water 2:274\u0026ndash;284. https://doi.org/10.1038/s44221-024-00210-z\u003c/li\u003e\n\u003cli\u003eZahura FT, Goodall JL (2022) Predicting combined tidal and pluvial flood inundation using a machine learning surrogate model. J Hydrol Reg Stud 41:101087. https://doi.org/10.1016/j.ejrh.2022.101087\u003c/li\u003e\n\u003cli\u003eZahura FT, Goodall JL, Sadler JM, et al (2020) Training Machine Learning Surrogate Models From a High-Fidelity Physics-Based Model: Application for Real-Time Street-Scale Flood Prediction in an Urban Coastal Community. Water Resour Res 56:e2019WR027038. https://doi.org/10.1029/2019WR027038\u003c/li\u003e\n\u003cli\u003eZhao G, Pang B, Xu Z, et al (2018) Mapping flood susceptibility in mountainous areas on a national scale in China. Sci Total Environ 615:1133\u0026ndash;1142. https://doi.org/10.1016/j.scitotenv.2017.10.037\u003c/li\u003e\n\u003cli\u003eZhao M, Liu Y, Wang Y, et al (2022) Effectiveness assessment of reservoir projects for flash flood control, water supply and irrigation in Wangmo Basin, China. Sci Total Environ 851:157918. https://doi.org/10.1016/j.scitotenv.2022.157918\u003c/li\u003e\n\u003cli\u003eZhou S, Jia W, Wang M, et al (2024) Synergistic assessment of multi-scenario urban waterlogging through data-driven decoupling analysis in high-density urban areas: A case study in Shenzhen, China. J Environ Manage 369:122330. https://doi.org/10.1016/j.jenvman.2024.122330\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"water-resources-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"warm","sideBox":"Learn more about [Water Resources Management](https://www.springer.com/journal/11269)","snPcode":"11269","submissionUrl":"https://submission.nature.com/new-submission/11269/3","title":"Water Resources Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Urban flood, Explainable machine learning, Flood inundation areas, Dominant driver","lastPublishedDoi":"10.21203/rs.3.rs-7476416/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7476416/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCompound floods, driven by concurrent heavy rainfall and high tides, increasingly threaten coastal urban areas with severe infrastructure damage and socioeconomic disruption. Rapid identification of inundation extents and their dominant drivers is crucial for enabling timely emergency response and infrastructure protection. Current machine learning (ML) models excel at predicting flood inundation extents, but their 'black-box' nature restricts identifying dominant local drivers. While explainable AI (XAI) techniques are emerging to address this challenge, it is necessary to determine how to best pair an XAI method with a high-performing ML model to ensure both predictive accuracy and robust interpretability. This study systematically evaluates Explainable Machine Learning (EML) models to predict inundation areas and identify dominant drivers. Our EML models were created by pairing two representative XAI techniques, SHAP and LIME, with three distinct types of ML models: a linear model (Logistic Regression), a tree-based ensemble (Random Forest), and a neural network (Multilayer Perceptron). Norfolk, Virginia, USA was selected to train, test, and conduct driver analysis for the models. Results revealed that the RF achieved the best predictive performance (Accuracy\u0026thinsp;=\u0026thinsp;0.81). Furthermore, of the two XAI techniques evaluated, SHAP-based driver attribution demonstrated greater consistency with real-world conditions because its ability to account for complex driver interactions, such as rainfall and tides, provides a more reliable identification of their influence. By leveraging XAI techniques, ML models can move beyond prediction to guiding informed decision-making and developing more effective flood management strategies.\u003c/p\u003e","manuscriptTitle":"Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-09 15:13:01","doi":"10.21203/rs.3.rs-7476416/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvited","content":"Water Resources Management","date":"2026-03-12T12:59:15+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-09-28T07:35:19+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-28T07:21:29+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-28T10:52:05+00:00","index":"","fulltext":""},{"type":"submitted","content":"Water Resources Management","date":"2025-08-28T01:21:14+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"water-resources-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"warm","sideBox":"Learn more about [Water Resources Management](https://www.springer.com/journal/11269)","snPcode":"11269","submissionUrl":"https://submission.nature.com/new-submission/11269/3","title":"Water Resources Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"5aea1179-95d7-4231-9fc3-71673f3d986b","owner":[],"postedDate":"October 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-10-09T15:13:01+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-09 15:13:01","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7476416","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7476416","identity":"rs-7476416","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}