AI-Powered Flood Monitoring for Azerbaijan Using Multi-Source Satellite Data: Operational Prototype Development and Initial Validation

Azerbaijan faces significant flood risks due to its diverse terrain, which includes low-lying areas near the Caspian Sea, situated 28 meters below sea level, and mountainous regions exceeding 4,000 meters. Increasing rainfall exacerbates this threat, while current monitoring systems are inadequate for timely flood warnings. This study introduces HydroAlert Azerbaijan, an artificial intelligence-driven system utilizing satellite imagery to detect floods and assess risk levels, capable of operating in adverse weather conditions. It employs a U-Net neural network to analyze Sentinel-1 SAR data, specifically leveraging VV and VH channels for efficient flood assessment.

The system processes 512×512 pixel tiles from the SAR data, overlapping by 64 pixels to ensure comprehensive coverage. Trained on the SEN12FLOOD dataset, consisting of 209 global flood examples, HydroAlert is designed to function effectively even in areas with limited flood event data. Initial evaluations of the Sentinel-1 SAR scenes and Azersky optical images confirm its efficacy, achieving an accuracy of approximately 85% in flood identification.
The Azersky optical data, characterized by a resolution of 1.5 meters, provides detailed insights into infrastructure vulnerability and validates the extent of floods derived from SAR data. The model generates precise vector shapes on maps, improving emergency response planning by visualizing flood extents.

This study shows that a platform facilitates user interaction with flood data, incorporating historical insights from the Dartmouth Flood Observatory and high-risk area alerts. The system supports data export in user-friendly formats to assist decision-making. The Hydroalert Project, which integrates SAR and optical data sources for comprehensive flood assessment, ensures reliable monitoring capabilities through its multi-sensor integration framework.

Additionally, the system incorporates a forecasting module using ConvLSTM architecture to predict flood risks over the following week, aiding proactive decision-making in disaster preparedness. Participation in the Azercosmos Earth Observation Competition 2025 has fostered collaboration with the Azerbaijani Space Agency, leading to systematic enhancements of the HydroAlert prototype using data from the competition.

Current efforts focus on refining the model for local conditions, utilizing satellite imagery to improve operational accuracy. This project demonstrates the potential of deep learning models for flood detection in developing regions lacking robust ground-level monitoring systems. By integrating global satellite images with advanced AI techniques, HydroAlert Azerbaijan offers a viable flood monitoring and management strategy for areas with limited existing resources and information.

 

Keywords

Flood mapping, SAR remote sensing, Optical imagery, Deep learning, Azerbaijan, Disaster management