ADS Capstone Chronicles Revised

18

typically used for image segmentation tasks. These models do not have a direct concept of "variable importance" like traditional ML models (e.g., random forests or decision tree). In CNNs like ResNet, MobileNet, and SimpleCNN, features importance doesn't exist because there are no discrete variables. Overall, this project succeeded in demonstrating and emphasizing the importance of applying automation to post-disaster damage assessments for an increase in the efficacy of recovery efforts. By providing timely and accurate impact reports, these models support first responders in prioritizing areas for intervention and ensuring optimal use of critical resources and infrastructure. Furthermore, the study shows the scalability of this approach, as it can be applied to a variety of disaster types, improving overall disaster management strategies, as each region throughout the nation and the world suffer from specific, contrasting, and detrimental natural disasters. The integration of ML with satellite imagery thus offers a powerful tool for transforming disaster response, ensuring faster recovery and more effective response across the globe. To enhance disaster response, it is recommended to expand the utilization of the xView2 geospatial imagery, which provides thousands of valuable pre- and post-disaster images. Having a limited file storage capacity, it’s also essential to optimize cloud storage and processing capabilities for quick access to imagery and model outputs. Real-world testing of the models in live disaster scenarios is crucial for fine-tuning and validating their effectiveness. Finally, developing intuitive dashboards from the developed model will enable first responders and decision-makers to easily access actionable insights in a digestible way to the users, and other stakeholders. 6.2 Recommended Action Items

6

Discussion 6.1 Conclusion

This project focused on enhancing disaster response efforts by integrating ML techniques with high-resolution satellite imagery for damage assessment and recovery operations. It highlighted the challenges of traditional disaster assessment methods, which often rely on ground-based or aerial surveys that are slow and resource intensive. By leveraging ML and satellite data, the project discussed automation solutions to streamline damage assessments to enable more accurate and fast disaster response, particularly applied in the aftermath of hurricanes, although can be demonstrated with any variety of natural disasters. The primary analytic objectives were threefold: optimize ML models, identify and localize building footprints in pre-disaster images, and classify damage levels in post-disaster images (e.g., no damage, minor damage, major damage, destroyed). The research employed datasets like the xView2, which included thousands of pre- and post-disaster satellite images, specifically focusing on hurricanes. The methodology involved the use of semantic segmentation models, such as U-Net and FCNs, for building localization, and CNNs like MobileNet for damage classification. These models showcased the segmenting of building regions in pre-disaster images and identified varying levels of damage in post-disaster images. The results yielded that U-Net performed exceptionally well in precision and recall, particularly for delineating building boundaries, while MobileNet outperformed other models in damage classification accuracy, offering a good balance between computational efficiency and classification robustness. These models were tested using various performance metrics such as precision, recall, F1-score, and IoU, which confirmed their capability to automate disaster assessments efficiently. FCNs, U-Net, SimpleCNN, ResNet, MobileNet models are

284