ADS Capstone Chronicles Revised

3

and Atmospheric Administration have primarily focused on ground and aerial assessments of disaster sites, with the limited use of satellite data and ML integration in a quick reaction response scenario (Amit & Aoki, 2017; Gupta et al., 2024; Hauptman et al., 2024). Comparing findings and gaps, such as limited ML applications with satellite data in real-time contexts, positions opportunities for advancing disaster response through improved integration of these technologies. One notable study from E3S Web of Conferences in 2024 focuses on damage level identification through MATLAB image processing techniques applied to GeoEye images. The study emphasizes that "The National Disaster Response Force (NDRF) and other disaster rescue organizations must use high-quality aerial photos taken by UAVs, or unmanned aerial vehicles, to create post-disaster strategies for recovering" (Gupta et al., 2024, p. 4). This highlights the direct importance of utilizing automation and high resolution data for a rapid response and an effective emergency recovery operation. Additionally, the study presents each method's benefits along with limitations, such as segmentation capability on isolating structural damage in images, or the precision of deep learning models in recognizing roads and construction features in these post catastrophe settings (Gupta et al., 2024). This research specifically analyzes regions affected by disasters in Thailand and Japan, demonstrating the effectiveness of using high-resolution satellite imagery to assess structural damage and inform recovery efforts. An emerging pattern observed in this study is the growing importance or desire for automated, high accuracy models like random forest, support vector machines, and convolutional neural networks (CNNs), which facilitate faster and more 3.1 A comparative analysis of post disaster analysis using image processing techniques

accurate damage assessments compared to traditional methods. Validated findings across multiple datasets indicate that ML models, especially when paired with high-resolution imagery, can achieve accuracies of up to 90% for specific types of damage detection, such as flooded regions and collapsed buildings. While many techniques show high accuracy, the results vary significantly depending on the type of disaster and its environment. For instance, bitemporal image classification methods work effectively in rural flooding scenarios but underperform in urban regions with dense infrastructure. Furthermore, there is limited research integrating satellite and social media data in real-time, a gap that, if addressed, could improve data synthesis and situational awareness in emergency response (Jayawardene et al., 2021). Another literature review from Doshi et al. (2018) employs semantic segmentation techniques for pixel-wise change detection in man-made features. This approach allows for a more granular understanding of how disasters affect urban landscapes, enabling quicker identification of affected infrastructure and facilitating targeted response efforts. The study follows a pattern that leverages CNNs and other ML models to improve the speed and accuracy of disaster assessments. Compared to traditional, manual disaster mapping, the Disaster Impact Index demonstrates high effectiveness and accuracy, with F1 scores of over 80% for both Hurricane Harvey and the Santa Rosa fire datasets. These findings validate the model's advantage in accurately pinpointing heavily impacted regions and offer strong evidence for CNNs’ reliability in high -stakes disaster contexts. However, the approach outlined in the article varies from the objectives of this project mainly on its focus of general impact detection rather than specific damage classification. The article emphasizes broad change detection by comparing 3.2 From satellite imagery to disaster insights

269