ADS Capstone Chronicles Revised

4

pre- and post-disaster satellite images to identify areas of significant impact, using general man made features as indicators of damage. Furthermore, in contrast, this project aims to classify damage levels with greater granularity such as no damage, minor, major, and destroyed. This detailed classification is intended to provide actionable insights, allowing response teams to allocate resources based on the severity of damage, rather than solely identifying areas of maximum impact. Additionally, this project includes the preliminary step of identifying precise building footprints in pre-disaster imagery through localization models, establishing a reliable baseline for post-disaster assessment — an aspect not prioritized in the article’s approa ch (Doshi et al., 2018). The field study held by International Electronics Symposium on Knowledge Creation and Intelligent Computing uses landslide and flood datasets alongside images sourced from Google Earth. By leveraging CNNs, this research demonstrates the potential of deep learning models to effectively classify and detect various types of disasters from aerial imagery. This study uses CNNs for automatic detection of these disasters by extracting features from “obtain[ing] two aerial imagery (pre-disaster and post-disaster aerial imagery for both landslide and flood) from Google Earth as input images” (Amit & Aoki, 2017, p. 2). While the article is showcasing the effectiveness of CNNs in detecting changes with the use of aerial imagery, it however is using traditional methods which comes with limitations which are predominantly relied on traditional sensor data and manual techniques. These such methods often suffered from limited range and accuracy due to the reliance on human interpretation and inadequate sensor coverage. By shifting the focus to automated CNN systems, the article 3.3 Disaster Detection from Aerial Imagery with Convolutional Neural Network

demonstrates a large significant advancement in disaster detection capabilities with the use of aerial imagery. However, this also highlights a gap in the exploration of various disaster types beyond floods and landslides, which could significantly improve the capability of the proposed methodology (Amit & Aoki, 2017).

3.4 Integrating ML and Remote Sensing in Disaster Management: A Decadal Review of Post-Disaster Building Damage Assessment

Another notable study where they use CNNs to develop rapidly detecting and categorizing structural damage. In hopes to further advance post- disaster recovery operations “researchers are employing various remote sensing technologies, including optical satellite imagery and synthetic aperture radar (SAR), to collect data on damaged buildings” (Al Shafian & Hu, 2024 , p. 11). This provides an in-depth understanding of how ML and remote sensing technologies have developed over the past few decades. It is highlighting the combination of satellite, unmanned aerial vehicle, and ground imagery paired with deep learning techniques to rapidly detect and categorize structural damage. Integration of multiple data sources can be very useful; however, it comes with significant challenges. Combining imagery from diverse sources often results in inconsistencies in format and resolution, as the data is originating from many different areas (i.e. unmanned aerial vehicle, satellites, ground). The data source synthesization portion is notably remarkable but can complicate the development of the models as it is not all completely harmonized data for something as critical as a real-world post disaster operation. Additionally, another gap that the research does not address is the limitations faced by ground-based observations which require significant resources and manpower (Al Shafian & Hu, 2024).

3.5 Hurricane Ian Damage Assessment Using Aerial Imagery and

270