ABS 021: From Reaction to Prevention: Predicting Flood Risk Using Pre-Disaster Satellite Imagery and Deep Learning

Dimitar Bogoeski ¹, Sai Korada ²

¹ University of Arizona
² University of California, Santa Cruz

The Van Wickle Journal (2026) Volume 2, ABS021

Introduction: Floods remain one of the most damaging natural hazards worldwide, and many current response systems focus more on reaction than prevention. This project explores whether deep learning can help identify flood-prone areas before a disaster occurs by using satellite imagery collected prior to flooding. Rather than relying only on post-disaster assessment, this work shifts attention toward early risk prediction, which could support better planning, resource allocation, and community preparedness. Using pre-disaster Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 optical satellite imagery, this study investigates whether meaningful surface patterns linked to future flood risk can be detected through a supervised convolutional neural network. By combining information from both radar and optical sensors, the model captures a wider range of environmental features than either source alone. This approach is especially useful because radar imagery can retain information under conditions where optical imagery may be limited, while optical imagery adds surface detail that may improve classification. The goal of this study was to develop a model that predicts flood risk from fused satellite image tiles and to evaluate its performance across multiple metrics. This work contributes to a growing body of research that uses artificial intelligence for disaster prevention and environmental risk analysis, while also showing how remote sensing and deep learning can be combined in a practical way for real-world decision-making.

Methods: A supervised convolutional neural network was developed to predict flood risk from pre-disaster satellite imagery. The model used fused, normalized 512×512 image tiles derived from Sentinel-1 SAR imagery and Sentinel-2 optical imagery. This multi-sensor approach was designed to combine the structural and surface-level information available across both data sources. The model was trained to classify flood risk using pre-disaster imagery only, with the goal of identifying patterns associated with future flooding before an event occurs. Model performance was evaluated using accuracy, ROC AUC, F1 score, precision, recall, and confusion matrices. In addition to standard evaluation metrics, the project also examined sensor-related and geographic bias to better understand how well the model performed across different conditions and settings.

Results: The model showed strong predictive performance in identifying flood risk from pre-disaster satellite imagery. It achieved 97.5% accuracy, a ROC AUC of 0.98, and an F1 score of 0.94. Evaluation through precision, recall, and confusion matrices supported the model’s ability to classify flood-prone areas effectively. These findings suggest that fused Sentinel-1 and Sentinel-2 imagery, when paired with a supervised CNN, can support accurate flood risk prediction before disaster onset.

Discussion: These findings suggest that deep learning with pre-disaster satellite imagery can support a more preventive approach to flood risk assessment. This matters because earlier identification of vulnerable areas could help communities and decision-makers act before damage occurs. The results also show the value of combining radar and optical data in one predictive framework. Future work could test the model across more geographic regions, improve generalizability, and explore how this approach could be used in real planning and emergency management systems.