GA-CDAM: Genetic Algorithm Driven Compact Damage Assessment Model Using Post-Disaster Aerial Images

Abstract

This article discusses the automated creation of a segmentation model from post-disaster aerial images for disaster analysis. Disaster damage assessment models must be efficient and accurate in order to provide relevant information on affected areas quickly. Manual design of such architectures is time-consuming and may not yield optimal results. To address this, we present a genetic algorithm that explores different components of UNet architectures to automatically design the best configuration. The goal is to maximize the accuracy of the segmentation model, which is the Mean Intersection over Union (MIoU) under explicit computational constraints. The algorithm explores a vast search space that contains various UNet architectural decisions (e.g., network depth, convolution type, oversampling strategies). It uses a fixed gene to generate a dynamic phenotype that serves as a neural network for the segmentation task. The results showed a significant reduction in model complexity, from over 40 million parameters for the State-Of-The-Art (SOTA) models to just 3.72 million parameters for our model. While maintaining good segmentation results, reaching 67.45% mIoU. We have successfully automated the design of the disaster damage assessment model by leveraging metaheuristic optimization and have provided a lightweight model that can be deployed in small devices for real-time disaster analysis.