Detection of Fire, non-fire and smoke using combination CNN and Transfer Learning

Detection of Fire, non-fire and smoke using combination CNN and Transfer 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 Detection of Fire, non-fire and smoke using combination CNN and Transfer Learning Shrikant Thakur, Pratham ., Sonia Arora, Pulkit Srivastava This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8651825/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract In the study, over Fire and Smoke Detection system by application of Convolutional Neural Networks (CNN) is presented by using a dataset comprising 3,500 images per class, namely fire, smoke, and non-fire, for training as well as for testing. The dataset was separated into training, validation, and testing subsets, with images being labelled as ‘fire’, ‘smoke’, or ‘non-fire’. A Convolutional Neural Networks (CNN) model utilizing a frozen ResNet50 architecture was trained using TensorFlow on Google Colab as well as on local system. The proposed model achieved an overall accuracy of 80.58% and an overall AUC score of 0.9429, illustrating a satisfactory generalization capability. The final trained model was evaluated on the test dataset. Additionally, the model was utilized for real time detection using Python and OpenCV, enabling the processing of camera inputs. At testing, the system successfully detected the events. In addition, the system can be configured to send alert notifications upon detection, making it suitable for early warning and prevention in domestic and industrial environments. Convolutional Neural Networks ResNet Transfer Learning Fire Detection Smoke Detection Computer Vision Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Detection of fire and smoke plays a critical role in preventing disasters and ensuring safety across diverse environments. The traditional detection systems majority of them uses sensor-based methods, which are often susceptible to false alarms and delayed responses. In response, deep learning approaches, particularly those using computer vision techniques, have shown improvements in identifying fire and smoke in real time. These models, especially Convolutional Neural Networks (CNNs) and some predefined libraries like resnet, giving a enhanced accuracy while minimizing false positives. However, training such models from start demands significant computational resources and access to large and a structured datasets an often-limiting factor. Transfer learning provides a promising solution by using pre-trained models like ResNet, VGG16, DenseNet121, allowing for the adaptation of knowledge from large-scale datasets for fire detection tasks with less data and training time. Yar et al. [ 1 ] conducted a broad comparison between machine learning methods for fire detection, underlining their limitations in accuracy and responsiveness. They also emphasized the superior performance of deep learning methods, particularly CNN-based models, in overcoming these challenges. In high-risk scenarios such as road tunnel fires, the need for advanced and robust detection systems becomes even more evident. Vision-based fire detection combined with vehicle tracking through deep learning not only enables real-time fire load estimation but also significantly contributes to emergency preparedness and tunnel safety management [ 4 ]. With the developing infrastructure of smart cities, integrating smart surveillance systems for fire detection has become more feasible. These systems use video for analysis supported by the machine learning approach, providing an enhanced accuracy in large, complicated environments where the traditional approaches often fail. The better results of deep learning techniques are still hindered by the scarcity of diverse and robust datasets. Moreover, vision-based models excel in giving precise results, they also encounter difficulties while distinguishing fire from objects having similar colour and detecting fires at some distance [ 10 ]. As a result, video-based fire detection has developed as a reliable alternative method some like colour segmentation, motion analysis, and deep learning to achieve better accuracy of detection. Despite of the various steps and methods applied various challenges are there, some like particularly due to varying lighting conditions, poor image quality, and the risk of false alarms. Furthermore, while CNN-based models provide descent results in controlled settings, their performance tends to decline when dealing with complex imagery, such as drone and satellite views, due to limitations in perspective and computational efficiency. This restricts their deployment on resource-constrained edge IoT devices. Another issue but crucial hurdle lies in the models’ tendency to treat entire frames as homogeneous segments, often looking over the localized fire regions—especially in the presence of heavy smoke or adverse weather. The dependency over the conventional and less diverse datasets also limits the scope of many systems, as a result gaps in accurately detecting fires across wilderness areas, dense urban landscapes, and smoke-dominant environments are observed. Additionally, transferring models across different geographical region remains problematic, as shifts in vegetation, building materials, and atmospheric conditions are there that often lead to a drop in detection performance. To enhance real-world and diverse applicability, new approaches must prioritize fine-grained object detection, the creation of richer datasets, and architectural optimizations that favour real-time responsiveness under operational constraints. 2. Method and Methodology 2.1 Dataset The dataset which we are to be used in our study is the FOREST_FIRE_SMOKE_AND_NON_FIRE_DATASET, which contains images in three classes which can be listed as following: 1. Fire: Images containing visible flames of environments in Fig. 2.1.1 2. Smoke: Images showing smoke without any flames in Fig. 2.1.2 3. Non-fire: Images of normal forest scenes without above two categories in Fig. 2.1.3 The above set of data was pre-divided into training and testing sets. The respective training dataset was used for model development, while the respective testing dataset was reserved for final evaluation. Each class contained approximately 3,500 images, providing a balanced dataset with a total of 10,500 images. 2.2 Data Pre processing 2.2.1 Data Organization A structured data frame was created containing image file paths and corresponding class labels The training dataset was split into 80% training and 20% validation sets respectively using stratified sampling to maintain a decent class distribution 2.2.2 Image Pre-processing All images were resized to 224×224 pixels to match the input as required by ResNet50 architecture Pixel values were to be normalized by rescaling it to a range [0,1] 2.2.3 Data Augmentation To increase the model generalization, the following augmentation techniques were used only to the training dataset: Rotation: Random rotations up to ± 30 degrees was applied Width and height shifts: Random shifts up to 20% in both directions was applied Shear transformations: Random shearing of up to 20% was done Zoom transformations: Random zooming of up to 20% was done Horizontal flipping was also applied Brightness variation: Random adjustments between 70% and 130% of original brightness was also applied 2.3 Model Architecture The model employed transfer learning namely ResNet50 as the base model: 1. Base Model : Pre-trained ResNet50 with ImageNet weights Input shape: 224×224×3 (RGB images) All layers frozen to preserve the learned features 2. Custom Classification Head : Global Average Pooling was used to reduce spatial dimensions Batch Normalization was also used for stabilized learning First layer containing 512 neurons, ReLU activation, and having L2 regularization (0.0001) First dropout with 0.5 rate Second layer containing 256 neurons, ReLU activation, and having L2 regularization (0.0001) Second dropout with 0.4 rate Third layer containing 128 neurons, ReLU activation, and having L2 regularization (0.0001) Third dropout with 0.3 rate Final output layer with 3 neurons and SoftMax as the activation function 2.4 Training Protocol The model was then trained using the following configuration: Optimizer : The learning rate of 0.0001 with Adam to avoid overfitting and have stable learning Loss Function : Categorical cross-entropy is used for multi-class classification and also because it compares predicted probability distribution with actual class labels Batch Size : 32 was used balanced memory efficiency and also for better model performance Maximum Epochs : 10 was selected to prevent overfitting and also to reduce computing time, with early stopping when the validation loss does not improve Regularization Techniques : Early stopping with a patience of 5 epochs indicating monitoring of validation loss Learning rate reduction with factor 0.2 when the model does not learn and improve up to 3 epochs Hardware : Google Colab with GPU acceleration and local system. 2.5 Evaluation Methodology The model performance was evaluated on basis of following: Overall Accuracy : Percentage of correctly classified images from the rest Accuracy = Correct Predictions/Total Predictions Precision : It is used because it shows how much predicted positives were actually correct, Precision = True Positives/ (True Positive + False Positives) Recall : It used because it gives how much actual true cases values were predicted Recall = True Positive / (True Positive + False Negative) F1-score : It is used because it balances precision and recall using harmonic mean. F1 = 2* ((Precision * Recall)/ (Precision + Recall)) ROC Curves : For each class using a one-vs-rest approach and also because it helps to visualize the model’s ability to distinguish between classes across different decision boundaries Area Under the Curve (AUC) : For each class and overall, as its area under ROC curve and helps to measure how well model ranks prediction rather then relying on some fixed threshold. 3. Results and outcome 3.1 Model Performance The model achieved the following performance metrics on the provided test dataset: ● Overall Accuracy of 80.58% ● Overall AUC (One-vs-Rest) of 0.9429 Class-specific performance: Class Precision Recall F1-score AUC Smoke 0.85 0.68 0.76 0.92 Fire 0.79 0.86 0.82 0.95 Non-fire 0.79 0.87 0.83 0.95 Average metrics: ● Macro Average: ▪ Precision = 0.81 ▪ Recall = 0.81 ▪ F1-score = 0.80 ● Weighted Average: ▪ Precision = 0.81 ▪ Recall = 0.81 ▪ F1-score = 0.80 3.2 ROC Curve Analysis The ROC curves demonstrate excellent distinguishing ability for all three classes: ● The "Fire" class shows a detection performance with AUC = 0.95 ● The "non-fire" class also shows classification capability with AUC = 0.95 ● The "Smoke" class shows performance with AUC = 0.92 The high AUC values across all classes indicate that the model effectively distinguishes in-between all three categories with less false positives and false negatives. The ROC is the abbreviation of Receivers- operating characteristic curve which shows model performance across all the threshold values and calculated by calculating true positive rate and false positive rate. 3.3 Class-specific Analysis ● Fire Detection: High recall (0.86) suggests the model has less chance to miss actual fire instances, making it suitable for safety-critical applications ● Smoke Detection: Highest precision (0.85) but lowest recall (0.68), indicates that the model is confident when it detects smoke, and it sometimes misses smoke instances ● Non-fire Classification: Highest recall (0.87) shows better ability to identify normal forest scenes 4. Conclusion The above study tries to show the effectiveness of transfer learning ResNet50 with CNN for forest fire and smoke detection. The model achieves a decent performance with 80.58% accuracy and decent AUC scores across all classes. The results suggest that deep learning approaches can provide reliable detection capabilities for early forest fire warning systems. The model shows better differentiating ability for all classes, with a AUC score of 0.9429 The class "Fire" promises better recall (0.86) which indicates a good sensitivity for fire detection The class "Smoke" also shows a precision (0.85) but having significantly low recall (0.68), providing an area for improvement Transfer learning with ResNet50 with CNN provides a foundation for fire detection tasks The detection method can be done in a better way by implementing class specific data set and by using better augmentation techniques. Also, some other pretrained models can be used to get refined outcome. For practical deployment, the model can be optimized for edge devices to ensure efficient performance in resource-constrained environments. Additionally, integrating it with existing forest monitoring systems would enhance its usability, while enabling real-time processing capabilities would make it more effective for rapid fire and smoke detection. Declarations Author Contribution S.T. (Shrikant Thakur) developed the model, collected the dataset from Kaggle, and wrote the manuscript. P. (Pratham) implemented the model as an application. S.A. (Sonia Arora) provided guidance on model generation, and P. (Pulkit) provided guidance on result analysis. All authors reviewed and approved the manuscript. Funding: The authors have not received any specific funding regarding this work. References Yar, H., Khan, Z. A., Rida, I., Ullah, W., Kim, M. J., & Baik, S. W. (2024). An efficient deep learning architecture for effective fire detection in smart surveillance. Image and Vision Computing, 145, 104989. https://doi.org/10.1016/j.imavis.2024.104989 Pincott, J., Tien, P. W., Wei, S., & Kaiser Calautit, J. (2022). Development and evaluation of a vision-based transfer learning approach for indoor fire and smoke detection. Building Services Engineering Research and Technology , 43 (3), 319-332. Reis, H. C., & Turk, V. (2023). Detection of forest fire using deep convolutional neural networks with transfer learning approach. Applied Soft Computing , 143 , 110362. Zhang, X., Chen, X., Ding, Y., Zhang, Y., Wang, Z., Shi, J., ... & Huang, X. (2024). Smart real-time evaluation of tunnel fire risk and evacuation safety via computer vision. Safety Science , 177 , 106563. Yandouzi, M., Grari, M., Idrissi, I., Boukabous, M., Moussaoui, O., Azizi, M., ... & Elmiad, A. K. (2022). Forest fires detection using deep transfer learning. Forest , 13 (8), 1. Sierra, D., Montanaro, W., Kuo, L., & Zohuri, B. (2023). Enhancing fire detection through CNN and transfer learning: A comprehensive research study. Journal of Engineering and Applied Sciences Technology. SRC/JEAST-242 , 174 (5), 2-6. Lv, H., & Chen, X. (2021, October). Research and implementation of forest fire smoke detection based on resnet transfer learning. In Proceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering (pp. 630-635). Sousa, M. J., Moutinho, A., & Almeida, M. (2020). Wildfire detection using transfer learning on augmented datasets. Expert Systems with Applications , 142 , 112975. Hong, Z., Hamdan, E., Zhao, Y., Ye, T., Pan, H., & Cetin, A. E. (2024). Wildfire detection via transfer learning: a survey. Signal, Image and Video Processing , 18 (1), 207-214. Sharma, H., & Kanwal, N. (2025). Intelligent video-based fire detection: A novel dataset and real-time multi-stage classification approach. Expert Systems with Applications , 126655. Dogan, S., Barua, P. D., Kutlu, H., Baygin, M., Fujita, H., Tuncer, T., & Acharya, U. R. (2022). Automated accurate fire detection system using ensemble pretrained residual network. Expert Systems with Applications , 203 , 117407. Dogan, S., Barua, P. D., Kutlu, H., Baygin, M., Fujita, H., Tuncer, T., & Acharya, U. R. (2022). Automated accurate fire detection system using ensemble pretrained residual network. Expert Systems with Applications , 203 , 117407. Jabnouni, H., Arfaoui, I., Cherni, M. A., Bouchouicha, M., & Sayadi, M. (2022, May). ResNet-50 based fire and smoke images classification. In 2022 6th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP) (pp. 1-6). IEEE. Bahhar, C., Ksibi, A., Ayadi, M., Jamjoom, M. M., Ullah, Z., Soufiene, B. O., & Sakli, H. (2023). Wildfire and smoke detection using staged YOLO model and ensemble CNN. Electronics , 12 (1), 228. Dubey, D., Chandra, S., Ghosh, D., Sinha, A. R., Ghose, C., Singh, J., ... & Bhattacharya, S. (2024, December). Fire Smoke Detection System using Machine Learning Models. In 2024 6th International Conference on Computational Intelligence and Networks (CINE) (pp. 1-6). IEEE. Geetha, S., Abhishek, C. S., & Akshayanat, C. S. (2021). Machine vision based fire detection techniques: A survey. Fire technology , 57 (2), 591-623. Khan, Z. A., Ullah, F. U. M., Yar, H., Ullah, W., Khan, N., Kim, M. J., & Baik, S. W. (2025). Optimized cross-module attention network and medium-scale dataset for effective fire detection. Pattern Recognition , 161 , 111273. Dalal, S., Lilhore, U. K., Radulescu, M., Simaiya, S., Jaglan, V., & Sharma, A. (2024). A hybrid LBP-CNN with YOLO-v5-based fire and smoke detection model in various environmental conditions for environmental sustainability in smart city. Environmental Science and Pollution Research , 1-18. Park, M., Tran, D. Q., Lee, S., & Park, S. (2021). Multilabel image classification with deep transfer learning for decision support on wildfire response. Remote Sensing , 13 (19), 3985. Taş, M., Taş, Y., Balki, O., Aydın, Z., & Taşdemir, K. (2022). Camera-based wildfire smoke detection for foggy environments. Journal of Electronic Imaging , 31 (5), 053033-053033. Tsalera, E., Papadakis, A., Voyiatzis, I., & Samarakou, M. (2023). CNN-based, contextualized, real-time fire detection in computational resource-constrained environments. Energy Reports , 9 , 247-257. Wang, Y., Wang, Y., Xu, C., Wang, X., & Zhang, Y. (2024). Computer vision-driven forest wildfire and smoke recognition via IoT drone cameras. Wireless Networks , 30 (9), 7603-7616. Marbach, G., Loepfe, M., & Brupbacher, T. (2006). An image processing technique for fire detection in video images. Fire safety journal , 41 (4), 285-289. Yunusov, N., Islam, B. M. S., Abdusalomov, A., & Kim, W. (2024). Robust Forest Fire Detection Method for Surveillance Systems Based on You Only Look Once Version 8 and Transfer Learning Approaches. Processes , 12 (5), 1039. Rajbhandari, A., Pokhrel, D., Gajurel, S., & Shakya, C. S. (2025). Performance Evaluations of the Deep Learning Models in Reference to Real-Time Fire and Smoke Detections Abilities. Kathford Journal of Engineering and Management , 4 (1), 73-83. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 22 Jan, 2026 Editor assigned by journal 22 Jan, 2026 Submission checks completed at journal 22 Jan, 2026 First submitted to journal 20 Jan, 2026 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-8651825","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":578270190,"identity":"1f47c381-6e62-4265-b938-5d9b35416a99","order_by":0,"name":"Shrikant Thakur","email":"data:image/png;base64,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","orcid":"","institution":"Dr. A.P.J. Abdul Kalam Technical University","correspondingAuthor":true,"prefix":"","firstName":"Shrikant","middleName":"","lastName":"Thakur","suffix":""},{"id":578270191,"identity":"adba3d90-ff49-446d-85aa-38f257ab6736","order_by":1,"name":"Pratham .","email":"","orcid":"","institution":"Dr. A.P.J. Abdul Kalam Technical University","correspondingAuthor":false,"prefix":"","firstName":"Pratham","middleName":"","lastName":".","suffix":""},{"id":578270192,"identity":"0c830c41-b78e-4f2b-bdac-95b9c7436235","order_by":2,"name":"Sonia Arora","email":"","orcid":"","institution":"Dr. A.P.J. Abdul Kalam Technical University","correspondingAuthor":false,"prefix":"","firstName":"Sonia","middleName":"","lastName":"Arora","suffix":""},{"id":578270193,"identity":"b32e37a5-eac2-407d-b6e2-1731520ad289","order_by":3,"name":"Pulkit Srivastava","email":"","orcid":"","institution":"Dr. A.P.J. Abdul Kalam Technical University","correspondingAuthor":false,"prefix":"","firstName":"Pulkit","middleName":"","lastName":"Srivastava","suffix":""}],"badges":[],"createdAt":"2026-01-20 17:55:00","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8651825/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8651825/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100877077,"identity":"f5a45280-42db-436b-8962-e01a0ece738b","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1335064,"visible":true,"origin":"","legend":"","description":"","filename":"ProjectPaper1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/bf551f3fa2f86184e3802a1b.docx"},{"id":100877074,"identity":"ef2214ef-5957-48b4-97ec-55d6c9e96825","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5360,"visible":true,"origin":"","legend":"","description":"","filename":"f55000e204d543a39003540c189e878f.json","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/3f913d63d45ce1b978f7d0f4.json"},{"id":100950786,"identity":"a3b1e541-1356-4db0-8a27-36a07994bd3a","added_by":"auto","created_at":"2026-01-23 07:09:12","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66304,"visible":true,"origin":"","legend":"","description":"","filename":"f55000e204d543a39003540c189e878f1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/65c51f1016bcbd65df3c2ff5.xml"},{"id":100877076,"identity":"6703fb06-d43d-4061-a28b-adc70ee34628","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":363909,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/741bc7338f6082169d707da7.jpeg"},{"id":100877079,"identity":"4ecd67f7-6f55-468d-bda3-99cfe0be3ed3","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":370423,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/52d84fc7b2602f567d6b980d.jpeg"},{"id":100877084,"identity":"926ea6d3-8cfb-4373-9881-847919c5fe4b","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":232054,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/8445ccbbd21317cac6754975.jpeg"},{"id":100877080,"identity":"19bc9ca3-6d85-4696-9441-2d328db4a41f","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":135786,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/2d1c8e94c6fe29270c6e265f.png"},{"id":100877082,"identity":"790e6ec5-3210-4950-ae2f-52a4ccd58bb8","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101624,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/9d992411e9bfa7a7e6892e10.png"},{"id":100877088,"identity":"012d4994-1875-45df-993f-d2b5ad81a6ca","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":166214,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/bde4240f9766499d5b52f664.png"},{"id":100877086,"identity":"957325d8-f9ca-46be-8dbe-278a8f2a1b87","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61362,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/deec233e8ba0b790c8b9f69d.png"},{"id":100877083,"identity":"501c8096-bcb8-423a-8042-5fb9e075b72c","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20581,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/238cb883a52f00116424d708.png"},{"id":100877089,"identity":"4687d82d-0b02-460a-8aaf-61cc40885205","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62256,"visible":true,"origin":"","legend":"","description":"","filename":"f55000e204d543a39003540c189e878f1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/3bad25366c6a53a78b3ba34d.xml"},{"id":100877087,"identity":"06fc1329-2843-41d7-8a91-f590d51ca5cb","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":76340,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/224d24d3bb7f28b40b69c9d0.html"},{"id":100950283,"identity":"8ae4457c-2eb5-4049-9010-6ccf963f970f","added_by":"auto","created_at":"2026-01-23 07:07:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":213031,"visible":true,"origin":"","legend":"\u003cp\u003eFire Images of Data Set\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/a8b04632023a821f044b55b0.png"},{"id":100877073,"identity":"c7b2e964-ce6c-4d7d-a054-f53ba001957b","added_by":"auto","created_at":"2026-01-22 10:27:46","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":211969,"visible":true,"origin":"","legend":"\u003cp\u003eNon-Fire Images of Dataset\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/2e77a436d74aa9d6c1466d04.png"},{"id":100877085,"identity":"44682b32-49c4-4f12-badb-bd34d496671b","added_by":"auto","created_at":"2026-01-22 10:27:47","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":136092,"visible":true,"origin":"","legend":"\u003cp\u003eSmoke Images of Dataset\u003c/p\u003e\n\u003cp\u003eSource:\u003ca href=\"https://www.kaggle.com/datasets/amerzishminha/forest-fire-smoke-and-non-fire-image-dataset\"\u003ehttps://www.kaggle.com/datasets/amerzishminha/forest-fire-smoke-and-non-fire-image-dataset\u003c/a\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/020d3653dd120f8beb55e903.png"},{"id":100950092,"identity":"aa13f76f-bbb7-48b3-a510-620997656c4f","added_by":"auto","created_at":"2026-01-23 07:06:51","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":66201,"visible":true,"origin":"","legend":"\u003cp\u003eROC Curve\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/a803cea23f13cf549ad91479.png"},{"id":100953114,"identity":"0a37213e-b0a5-4b12-b100-f8307d9db937","added_by":"auto","created_at":"2026-01-23 07:20:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1351814,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8651825/v1/4d44263a-69ed-481b-96f5-164bcfa96b8f.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Detection of Fire, non-fire and smoke using combination CNN and Transfer Learning","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eDetection of fire and smoke plays a critical role in preventing disasters and ensuring safety across diverse environments. The traditional detection systems majority of them uses sensor-based methods, which are often susceptible to false alarms and delayed responses. In response, deep learning approaches, particularly those using computer vision techniques, have shown improvements in identifying fire and smoke in real time. These models, especially Convolutional Neural Networks (CNNs) and some predefined libraries like resnet, giving a enhanced accuracy while minimizing false positives. However, training such models from start demands significant computational resources and access to large and a structured datasets an often-limiting factor. Transfer learning provides a promising solution by using pre-trained models like ResNet, VGG16, DenseNet121, allowing for the adaptation of knowledge from large-scale datasets for fire detection tasks with less data and training time.\u003c/p\u003e \u003cp\u003eYar et al. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] conducted a broad comparison between machine learning methods for fire detection, underlining their limitations in accuracy and responsiveness. They also emphasized the superior performance of deep learning methods, particularly CNN-based models, in overcoming these challenges. In high-risk scenarios such as road tunnel fires, the need for advanced and robust detection systems becomes even more evident. Vision-based fire detection combined with vehicle tracking through deep learning not only enables real-time fire load estimation but also significantly contributes to emergency preparedness and tunnel safety management [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWith the developing infrastructure of smart cities, integrating smart surveillance systems for fire detection has become more feasible. These systems use video for analysis supported by the machine learning approach, providing an enhanced accuracy in large, complicated environments where the traditional approaches often fail. The better results of deep learning techniques are still hindered by the scarcity of diverse and robust datasets. Moreover, vision-based models excel in giving precise results, they also encounter difficulties while distinguishing fire from objects having similar colour and detecting fires at some distance [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAs a result, video-based fire detection has developed as a reliable alternative method some like colour segmentation, motion analysis, and deep learning to achieve better accuracy of detection. Despite of the various steps and methods applied various challenges are there, some like particularly due to varying lighting conditions, poor image quality, and the risk of false alarms. Furthermore, while CNN-based models provide descent results in controlled settings, their performance tends to decline when dealing with complex imagery, such as drone and satellite views, due to limitations in perspective and computational efficiency. This restricts their deployment on resource-constrained edge IoT devices. Another issue but crucial hurdle lies in the models\u0026rsquo; tendency to treat entire frames as homogeneous segments, often looking over the localized fire regions\u0026mdash;especially in the presence of heavy smoke or adverse weather. The dependency over the conventional and less diverse datasets also limits the scope of many systems, as a result gaps in accurately detecting fires across wilderness areas, dense urban landscapes, and smoke-dominant environments are observed. Additionally, transferring models across different geographical region remains problematic, as shifts in vegetation, building materials, and atmospheric conditions are there that often lead to a drop in detection performance. To enhance real-world and diverse applicability, new approaches must prioritize fine-grained object detection, the creation of richer datasets, and architectural optimizations that favour real-time responsiveness under operational constraints.\u003c/p\u003e"},{"header":"2. Method and Methodology","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1 Dataset\u003c/h2\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe dataset which we are to be used in our study is the FOREST_FIRE_SMOKE_AND_NON_FIRE_DATASET, which contains images in three classes which can be listed as following:\u003c/p\u003e\n \u003c/div\u003e\u003cspan\u003e\n \u003cp\u003e1. Fire: Images containing visible flames of environments in Fig. 2.1.1\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Smoke: Images showing smoke without any flames in Fig. 2.1.2\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Non-fire: Images of normal forest scenes without above two categories in Fig. 2.1.3\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe above set of data was pre-divided into training and testing sets. The respective training dataset was used for model development, while the respective testing dataset was reserved for final evaluation. Each class contained approximately 3,500 images, providing a balanced dataset with a total of 10,500 images.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2 Data Pre processing\u003c/h2\u003e\n \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e\n \u003ch2\u003e2.2.1 Data Organization\u003c/h2\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eA structured data frame was created containing image file paths and corresponding class labels\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eThe training dataset was split into 80% training and 20% validation sets respectively using stratified sampling to maintain a decent class distribution\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\n \u003ch2\u003e2.2.2 Image Pre-processing\u003c/h2\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eAll images were resized to 224\u0026times;224 pixels to match the input as required by ResNet50 architecture\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003ePixel values were to be normalized by rescaling it to a range [0,1]\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\n \u003ch2\u003e2.2.3 Data Augmentation\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eTo increase the model generalization, the following augmentation techniques were used only to the training dataset:\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eRotation: Random rotations up to \u0026plusmn;\u0026thinsp;30 degrees was applied\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eWidth and height shifts: Random shifts up to 20% in both directions was applied\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eShear transformations: Random shearing of up to 20% was done\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eZoom transformations: Random zooming of up to 20% was done\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eHorizontal flipping was also applied\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eBrightness variation: Random adjustments between 70% and 130% of original brightness was also applied\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3 Model Architecture\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe model employed transfer learning namely ResNet50 as the base model:\u003c/p\u003e\n \u003c/div\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e1. Base Model\u003c/strong\u003e: Pre-trained ResNet50 with ImageNet weights\u003c/p\u003e\n \u003c/span\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eInput shape: 224\u0026times;224\u0026times;3 (RGB images)\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eAll layers frozen to preserve the learned features\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e2. Custom Classification Head\u003c/strong\u003e:\u003c/p\u003e\n \u003c/span\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eGlobal Average Pooling was used to reduce spatial dimensions\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eBatch Normalization was also used for stabilized learning\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eFirst layer containing 512 neurons, ReLU activation, and having L2 regularization (0.0001)\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eFirst dropout with 0.5 rate\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eSecond layer containing 256 neurons, ReLU activation, and having L2 regularization (0.0001)\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eSecond dropout with 0.4 rate\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eThird layer containing 128 neurons, ReLU activation, and having L2 regularization (0.0001)\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eThird dropout with 0.3 rate\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eFinal output layer with 3 neurons and SoftMax as the activation function\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 Training Protocol\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe model was then trained using the following configuration:\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eOptimizer\u003c/strong\u003e: The learning rate of 0.0001 with Adam to avoid overfitting and have stable learning\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eLoss Function\u003c/strong\u003e: Categorical cross-entropy is used for multi-class classification and also because it compares predicted probability distribution with actual class labels\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eBatch Size\u003c/strong\u003e: 32 was used balanced memory efficiency and also for better model performance\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eMaximum Epochs\u003c/strong\u003e: 10 was selected to prevent overfitting and also to reduce computing time, with early stopping when the validation loss does not improve\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eRegularization Techniques\u003c/strong\u003e:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eEarly stopping with a patience of 5 epochs indicating monitoring of validation loss\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eLearning rate reduction with factor 0.2 when the model does not learn and improve up to 3 epochs\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eHardware\u003c/strong\u003e: Google Colab with GPU acceleration and local system.\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e2.5 Evaluation Methodology\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe model performance was evaluated on basis of following:\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eOverall Accuracy\u003c/strong\u003e: Percentage of correctly classified images from the rest\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u0026thinsp;=\u003c/strong\u003e\u0026thinsp;Correct Predictions/Total Predictions\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e: It is used because it shows how much predicted positives were actually correct,\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u0026thinsp;=\u003c/strong\u003e\u0026thinsp;True Positives/ (True Positive\u0026thinsp;+\u0026thinsp;False Positives)\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e: It used because it gives how much actual true cases values were predicted\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u0026thinsp;=\u003c/strong\u003e\u0026thinsp;True Positive / (True Positive\u0026thinsp;+\u0026thinsp;False Negative)\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eF1-score\u003c/strong\u003e: It is used because it balances precision and recall using harmonic mean.\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1\u0026thinsp;=\u003c/strong\u003e\u0026thinsp;2* ((Precision * Recall)/ (Precision\u0026thinsp;+\u0026thinsp;Recall))\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eROC Curves\u003c/strong\u003e: For each class using a one-vs-rest approach and also because it helps to visualize the model\u0026rsquo;s ability to distinguish between classes across different decision boundaries\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eArea Under the Curve (AUC)\u003c/strong\u003e: For each class and overall, as its area under ROC curve and helps to measure how well model ranks prediction rather then relying on some fixed threshold.\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e"},{"header":"3. Results and outcome","content":"\u003cp\u003e\u003cstrong\u003e3.1 Model Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe model achieved the following performance metrics on the provided test dataset:\u003c/p\u003e\n\u003cp\u003e● Overall Accuracy of 80.58%\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e● Overall AUC (One-vs-Rest) of 0.9429\u003c/p\u003e\n\u003cp\u003eClass-specific performance:\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eClass\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eSmoke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eFire\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eNon-fire\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAverage metrics:\u003c/p\u003e\n\u003cp\u003e● Macro Average:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e▪ Precision = 0.81\u003c/p\u003e\n\u003cp\u003e▪ \u0026nbsp;Recall = 0.81\u003c/p\u003e\n\u003cp\u003e▪ \u0026nbsp;F1-score = 0.80\u003c/p\u003e\n\u003cp\u003e● Weighted Average:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e▪ Precision = 0.81\u003c/p\u003e\n\u003cp\u003e▪ \u0026nbsp;Recall = 0.81\u003c/p\u003e\n\u003cp\u003e▪ \u0026nbsp;F1-score = 0.80\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 ROC Curve Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ROC curves demonstrate excellent distinguishing ability for all three classes:\u003c/p\u003e\n\u003cp\u003e● The \u0026quot;Fire\u0026quot; class shows a detection performance with AUC = 0.95\u003c/p\u003e\n\u003cp\u003e● The \u0026quot;non-fire\u0026quot; class also shows classification capability with AUC = 0.95\u003c/p\u003e\n\u003cp\u003e● The \u0026quot;Smoke\u0026quot; class shows performance with AUC = 0.92\u003c/p\u003e\n\u003cp\u003eThe high AUC values across all classes indicate that the model effectively distinguishes in-between all three categories with less false positives and false negatives. The ROC is the abbreviation of Receivers- operating characteristic curve which shows model performance across all the threshold values and calculated by calculating true positive rate and false positive rate.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Class-specific Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e● Fire Detection: High recall (0.86) suggests the model has less chance to miss actual fire instances, making it suitable for safety-critical applications\u003c/p\u003e\n\u003cp\u003e● Smoke Detection: Highest precision (0.85) but lowest recall (0.68), indicates that the model is confident when it detects smoke, and it sometimes misses smoke instances\u003c/p\u003e\n\u003cp\u003e● Non-fire Classification: Highest recall (0.87) shows better ability to identify normal forest scenes\u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe above study tries to show the effectiveness of transfer learning ResNet50 with CNN for forest fire and smoke detection. The model achieves a decent performance with 80.58% accuracy and decent AUC scores across all classes. The results suggest that deep learning approaches can provide reliable detection capabilities for early forest fire warning systems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eThe model shows better differentiating ability for all classes, with a AUC score of 0.9429\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe class \"Fire\" promises better recall (0.86) which indicates a good sensitivity for fire detection\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe class \"Smoke\" also shows a precision (0.85) but having significantly low recall (0.68), providing an area for improvement\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTransfer learning with ResNet50 with CNN provides a foundation for fire detection tasks\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe detection method can be done in a better way by implementing class specific data set and by using better augmentation techniques. Also, some other pretrained models can be used to get refined outcome. For practical deployment, the model can be optimized for edge devices to ensure efficient performance in resource-constrained environments. Additionally, integrating it with existing forest monitoring systems would enhance its usability, while enabling real-time processing capabilities would make it more effective for rapid fire and smoke detection.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eS.T. (Shrikant Thakur) developed the model, collected the dataset from Kaggle, and wrote the manuscript. P. (Pratham) implemented the model as an application. S.A. (Sonia Arora) provided guidance on model generation, and P. (Pulkit) provided guidance on result analysis. All authors reviewed and approved the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have not received any specific funding regarding this work.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eYar, H., Khan, Z. A., Rida, I., Ullah, W., Kim, M. J., \u0026amp; Baik, S. W. (2024). An efficient deep learning architecture for effective fire detection in smart surveillance. Image and Vision Computing, 145, 104989. https://doi.org/10.1016/j.imavis.2024.104989 \u003c/li\u003e\n\u003cli\u003ePincott, J., Tien, P. W., Wei, S., \u0026amp; Kaiser Calautit, J. (2022). Development and evaluation of a vision-based transfer learning approach for indoor fire and smoke detection. \u003cem\u003eBuilding Services Engineering Research and Technology\u003c/em\u003e, \u003cem\u003e43\u003c/em\u003e(3), 319-332.\u003c/li\u003e\n\u003cli\u003eReis, H. C., \u0026amp; Turk, V. (2023). Detection of forest fire using deep convolutional neural networks with transfer learning approach. \u003cem\u003eApplied Soft Computing\u003c/em\u003e, \u003cem\u003e143\u003c/em\u003e, 110362.\u003c/li\u003e\n\u003cli\u003eZhang, X., Chen, X., Ding, Y., Zhang, Y., Wang, Z., Shi, J., ... \u0026amp; Huang, X. (2024). Smart real-time evaluation of tunnel fire risk and evacuation safety via computer vision. \u003cem\u003eSafety Science\u003c/em\u003e, \u003cem\u003e177\u003c/em\u003e, 106563.\u003c/li\u003e\n\u003cli\u003eYandouzi, M., Grari, M., Idrissi, I., Boukabous, M., Moussaoui, O., Azizi, M., ... \u0026amp; Elmiad, A. K. (2022). Forest fires detection using deep transfer learning. \u003cem\u003eForest\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(8), 1.\u003c/li\u003e\n\u003cli\u003eSierra, D., Montanaro, W., Kuo, L., \u0026amp; Zohuri, B. (2023). Enhancing fire detection through CNN and transfer learning: A comprehensive research study. \u003cem\u003eJournal of Engineering and Applied Sciences Technology. SRC/JEAST-242\u003c/em\u003e, \u003cem\u003e174\u003c/em\u003e(5), 2-6.\u003c/li\u003e\n\u003cli\u003eLv, H., \u0026amp; Chen, X. (2021, October). Research and implementation of forest fire smoke detection based on resnet transfer learning. In \u003cem\u003eProceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering\u003c/em\u003e (pp. 630-635).\u003c/li\u003e\n\u003cli\u003eSousa, M. J., Moutinho, A., \u0026amp; Almeida, M. (2020). Wildfire detection using transfer learning on augmented datasets. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e142\u003c/em\u003e, 112975.\u003c/li\u003e\n\u003cli\u003eHong, Z., Hamdan, E., Zhao, Y., Ye, T., Pan, H., \u0026amp; Cetin, A. E. (2024). Wildfire detection via transfer learning: a survey. \u003cem\u003eSignal, Image and Video Processing\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(1), 207-214.\u003c/li\u003e\n\u003cli\u003eSharma, H., \u0026amp; Kanwal, N. (2025). Intelligent video-based fire detection: A novel dataset and real-time multi-stage classification approach. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, 126655. \u003c/li\u003e\n\u003cli\u003eDogan, S., Barua, P. D., Kutlu, H., Baygin, M., Fujita, H., Tuncer, T., \u0026amp; Acharya, U. R. (2022). Automated accurate fire detection system using ensemble pretrained residual network. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e203\u003c/em\u003e, 117407.\u003c/li\u003e\n\u003cli\u003eDogan, S., Barua, P. D., Kutlu, H., Baygin, M., Fujita, H., Tuncer, T., \u0026amp; Acharya, U. R. (2022). Automated accurate fire detection system using ensemble pretrained residual network. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e203\u003c/em\u003e, 117407.\u003c/li\u003e\n\u003cli\u003eJabnouni, H., Arfaoui, I., Cherni, M. A., Bouchouicha, M., \u0026amp; Sayadi, M. (2022, May). ResNet-50 based fire and smoke images classification. In \u003cem\u003e2022 6th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eBahhar, C., Ksibi, A., Ayadi, M., Jamjoom, M. M., Ullah, Z., Soufiene, B. O., \u0026amp; Sakli, H. (2023). Wildfire and smoke detection using staged YOLO model and ensemble CNN. \u003cem\u003eElectronics\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(1), 228.\u003c/li\u003e\n\u003cli\u003eDubey, D., Chandra, S., Ghosh, D., Sinha, A. R., Ghose, C., Singh, J., ... \u0026amp; Bhattacharya, S. (2024, December). Fire Smoke Detection System using Machine Learning Models. In \u003cem\u003e2024 6th International Conference on Computational Intelligence and Networks (CINE)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eGeetha, S., Abhishek, C. S., \u0026amp; Akshayanat, C. S. (2021). Machine vision based fire detection techniques: A survey. \u003cem\u003eFire technology\u003c/em\u003e, \u003cem\u003e57\u003c/em\u003e(2), 591-623.\u003c/li\u003e\n\u003cli\u003eKhan, Z. A., Ullah, F. U. M., Yar, H., Ullah, W., Khan, N., Kim, M. J., \u0026amp; Baik, S. W. (2025). Optimized cross-module attention network and medium-scale dataset for effective fire detection. \u003cem\u003ePattern Recognition\u003c/em\u003e, \u003cem\u003e161\u003c/em\u003e, 111273.\u003c/li\u003e\n\u003cli\u003eDalal, S., Lilhore, U. K., Radulescu, M., Simaiya, S., Jaglan, V., \u0026amp; Sharma, A. (2024). A hybrid LBP-CNN with YOLO-v5-based fire and smoke detection model in various environmental conditions for environmental sustainability in smart city. \u003cem\u003eEnvironmental Science and Pollution Research\u003c/em\u003e, 1-18.\u003c/li\u003e\n\u003cli\u003ePark, M., Tran, D. Q., Lee, S., \u0026amp; Park, S. (2021). Multilabel image classification with deep transfer learning for decision support on wildfire response. \u003cem\u003eRemote Sensing\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(19), 3985.\u003c/li\u003e\n\u003cli\u003eTaş, M., Taş, Y., Balki, O., Aydın, Z., \u0026amp; Taşdemir, K. (2022). Camera-based wildfire smoke detection for foggy environments. \u003cem\u003eJournal of Electronic Imaging\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(5), 053033-053033.\u003c/li\u003e\n\u003cli\u003eTsalera, E., Papadakis, A., Voyiatzis, I., \u0026amp; Samarakou, M. (2023). CNN-based, contextualized, real-time fire detection in computational resource-constrained environments. \u003cem\u003eEnergy Reports\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 247-257.\u003c/li\u003e\n\u003cli\u003eWang, Y., Wang, Y., Xu, C., Wang, X., \u0026amp; Zhang, Y. (2024). Computer vision-driven forest wildfire and smoke recognition via IoT drone cameras. \u003cem\u003eWireless Networks\u003c/em\u003e, \u003cem\u003e30\u003c/em\u003e(9), 7603-7616.\u003c/li\u003e\n\u003cli\u003eMarbach, G., Loepfe, M., \u0026amp; Brupbacher, T. (2006). An image processing technique for fire detection in video images. \u003cem\u003eFire safety journal\u003c/em\u003e, \u003cem\u003e41\u003c/em\u003e(4), 285-289.\u003c/li\u003e\n\u003cli\u003eYunusov, N., Islam, B. M. S., Abdusalomov, A., \u0026amp; Kim, W. (2024). Robust Forest Fire Detection Method for Surveillance Systems Based on You Only Look Once Version 8 and Transfer Learning Approaches. \u003cem\u003eProcesses\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(5), 1039.\u003c/li\u003e\n\u003cli\u003eRajbhandari, A., Pokhrel, D., Gajurel, S., \u0026amp; Shakya, C. S. (2025). Performance Evaluations of the Deep Learning Models in Reference to Real-Time Fire and Smoke Detections Abilities. \u003cem\u003eKathford Journal of Engineering and Management\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(1), 73-83.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"signal-image-and-video-processing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"sivp","sideBox":"Learn more about [Signal, Image and Video Processing](http://link.springer.com/journal/11760)","snPcode":"11760","submissionUrl":"https://submission.nature.com/new-submission/11760/3","title":"Signal, Image and Video Processing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Convolutional Neural Networks, ResNet, Transfer Learning, Fire Detection, Smoke Detection, Computer Vision","lastPublishedDoi":"10.21203/rs.3.rs-8651825/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8651825/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn the study, over Fire and Smoke Detection system by application of Convolutional Neural Networks (CNN) is presented by using a dataset comprising 3,500 images per class, namely fire, smoke, and non-fire, for training as well as for testing. The dataset was separated into training, validation, and testing subsets, with images being labelled as \u0026lsquo;fire\u0026rsquo;, \u0026lsquo;smoke\u0026rsquo;, or \u0026lsquo;non-fire\u0026rsquo;. A Convolutional Neural Networks (CNN) model utilizing a frozen ResNet50 architecture was trained using TensorFlow on Google Colab as well as on local system. The proposed model achieved an overall accuracy of 80.58% and an overall AUC score of 0.9429, illustrating a satisfactory generalization capability. The final trained model was evaluated on the test dataset. Additionally, the model was utilized for real time detection using Python and OpenCV, enabling the processing of camera inputs. At testing, the system successfully detected the events. In addition, the system can be configured to send alert notifications upon detection, making it suitable for early warning and prevention in domestic and industrial environments.\u003c/p\u003e","manuscriptTitle":"Detection of Fire, non-fire and smoke using combination CNN and Transfer Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-22 10:27:42","doi":"10.21203/rs.3.rs-8651825/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-22T15:15:17+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-22T13:37:47+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-22T13:35:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"Signal, Image and Video Processing","date":"2026-01-20T16:58:33+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"signal-image-and-video-processing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"sivp","sideBox":"Learn more about [Signal, Image and Video Processing](http://link.springer.com/journal/11760)","snPcode":"11760","submissionUrl":"https://submission.nature.com/new-submission/11760/3","title":"Signal, Image and Video Processing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"d01d92de-5b1a-45f2-a6d0-7514b0185dcd","owner":[],"postedDate":"January 22nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-01-23T12:24:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-22 10:27:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8651825","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8651825","identity":"rs-8651825","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}