ADS Capstone Chronicles Revised

5

due to varying update times, ultimately affecting damage cost assessments

Light Detection and Ranging: A Case Study of Estero Island, Florida

On September 28, 2022, a category 4 catastrophic event otherwise known as Hurricane Ian made landfall in Lee County, Florida. An emergency initiative implemented defined as “the Coastal Resilience Task Force, [which is] a combination of local, state and federal agencies, is focused on the recovery of impacted coastal areas, restoration of living shorelines and other nature-based solutions to mitigate future coastal flooding and storm surge impacts” ( Federal Emergency Management Agency, 2023). This study investigates the advantageous nature of combining aerial imagery and light detection and ranging for conducting post-hurricane structural damage assessments. Something specific this article is doing is that it analyzes both structural damages and changes in beach morphology on Estero Island, Florida. Effectively using high-resolution imagery from National Oceanic and Atmospheric Administration and airborne light detection and ranging data, the researchers were able to map debris, assess building damages, and analyze morphological shifts. They developed a classification system based on Federal Emergency Management Agency guidelines, categorizing structures into damage levels from “not affected” to “destroyed,” which helped in identifying priority areas for immediate post-storm response and recovery planning. Furthermore, the study suffered from the limitations associated with aerial imagery and light detection and ranging which was that the methods “cannot accurately identify damages caused by storm surges and flooding” (Hauptman et al., 2024, p. 13). First responders would quickly discover that this “study may have underestimated the extent of damage to a structure” (Hauptman et al., 2024 , p. 13). Lastly, another gap is in property value estimation; the study notes that geographic information system based data may lead to under- or overestimations

4 Methodology

This project involves two core tasks: building localization and damage classification. The building localization task leverages a semantic segmentation model to detect and mask building footprints in pre-disaster satellite images. By taking pre-disaster images as input, the model outputs binary masks that delineate building regions, establishing areas of interest for subsequent analysis. The damage classification task employs a multiclass segmentation model to determine the level of damage for each pixel within the localized building regions in post disaster images. Trained on paired pre- and post disaster imagery, this model performs pixel-wise classification into four categories: undamaged, minor damage, major damage, and destroyed. All code used for data preparation, exploration, and modeling can be found at the following GitHub repository https://github.com/jeremiahf24/ADS Capstone.git The data was sourced from xView2, which provides label files in JSON format and images in PNG format. xView2 contains thousands of data entries covering various natural disaster types, including volcanoes, tsunamis, floods, fires, and hurricanes totaling 11,196 JSON and PNG files. However, for this project the focus is specifically on 4,876 pre- and post-hurricane JSON and image files, which are stored in an S3 bucket. The JSON and PNG files were exceedingly large in size and needed to be easily accessible to a geographically separated set of users. To meet these requirements, the team used Amazon’s S3 bucket for storage due to its scalability, durability, and ability to provide seamless controlled access for these sizable files (Amazon Web Services, n.d.). 4.1 Data Acquisition and Aggregation

271