ADS Capstone Chronicles Revised

2

responders in damaged or uncertain zones is paramount. Integrating ML capabilities with satellite imagery presents a transformative approach to overcoming these limitations in disaster management. 2.1 Problem Identification and Motivation Natural disasters consistently challenge traditional disaster response methods, underscoring the need for optimized, timely, and accurate impact assessments. Leveraging ML models on satellite imagery provides actionable insights to first responders, enhancing humanitarian aid efforts and potentially saving lives by accelerating the allocation of critical resources. ● Identifying building footprints within pre-disaster satellite images using localization models. ● Classifying damage levels (no damage, minor, major, or destroyed) in post disaster images to aid in efficient response. by leveraging a multi-metric evaluation of the xView2 dataset to ensure accuracy, robustness, and generalizability across the diverse disaster scenarios. 3 Literature Review (related works) The literature review effectively positions this project within current research on satellite imagery and ML for disaster response, specifically focusing on post-disaster damage assessments. After analyzing prior studies, this review will identify key themes, recurring patterns, and validated findings, as well as note contrasting approaches and any identified gaps. Previous studies conducted by the International Electronics Symposium on Knowledge Creation and Intelligent Computing and the National Oceanic ● Optimizing model performance 2.2 Definition of Objectives The project’s objectives include:

KEYWORDS Natural Disaster, Machine Learning, Satellite Imagery, Damage Assessment, Humanitarian Response, xView2 Dataset 1 Introduction Natural disasters can occur in a variety of environments and communities and can strike unpredictably. This fact is true of hurricanes, wildfires, and earthquakes which can significantly disrupt communities on a large scale likely without warning. Traditional aerial-based and ground-based damage assessment methods currently used in today ’s society , while effective, are frequently hindered by damaged roadways, poor weather, and hazardous conditions. This application can be used across various natural disasters — such as wildfires, tsunamis, and earthquakes — to analyze pre- and post-disaster imagery for efficient triage and response. However, this analysis will primarily focus on hurricanes, specifically in response to the recent devastation caused by Hurricane Helene, which impacted millions in the southeastern United States. To overcome these limitations, this project proposes the use of machine learning (ML) models integrated into geospatial analysis of satellite and remote sensing imagery to automate damage assessments and improve their efficacy during dire situations. This approach aims to speed up first responder support and optimize resource allocation for recovery efforts. 2 Background Traditional damage assessment techniques demand significant personnel, time, and resources, which can hinder response efficiency. Historical examples, such as Hurricane Katrina in 2005, Hurricane Michael in 2018, and most recently Hurricane Helene in 2024, underscore the ongoing challenges of disaster recovery. Often these events are earmarked by significantly delayed response times, limited resource allocation and distribution, and multiple casualties or fatalities throughout affected areas. Most critically, the safety of first

268