DisasterVision AI: A Satellite-Based Deep Learning Framework for Automated Disaster Damage Assessment

Natural disasters such as earthquakes, floods, hurricanes, and wildfires cause extensive damage to infrastructure and human life, making rapid and accurate damage assessment a critical requirement for effective disaster response. Traditional ground-based assessment techniques are time-consuming, risky, and limited in spatial coverage, which delays emergency decision-making processes. To address these limitations, this paper presents DisasterVision AI, an automated satellite imagery analysis system that leverages deep learning for large-scale building damage assessment. The proposed system utilizes a modified Single Shot MultiBox Detector (SSD) with a VGG-16 backbone, enhanced to process six-channel input by combining pre-disaster and post-disaster satellite images. This dualinput architecture enables the model to learn visual differences between temporal image pairs, improving damage detection accuracy. The model is trained using the xView2 dataset, which provides annotated satellite imagery with four damage categories: no-damage, minor-damage, major-damage, and destroyed. The system incorporates advanced training techniques including data augmentation using Albumentations, OneCycle learning rate scheduling, and AdamW optimization for efficient convergence. Performance evaluation is conducted using Mean Average Precision (mAP) metrics across multiple IoU thresholds. Additionally, Non-Maximum Suppression (NMS) is applied for refining detection outputs. Experimental results demonstrate that DisasterVision AI provides fast, scalable, and reliable damage assessment, making it a valuable tool for disaster management authorities and emergency response teams.