STTORM-CD: Low-Demand and High-Impact Disaster Monitoring Onboard Satellites Using Change Detection

Satellite imagery can play a crucial role in disaster management. However, without extensive costs, critical images may take hours or even days to reach end-users. This article explores change detection methods for real-time disaster identification onboard satellites as an alternative method to decrease reaction time. We introduce STTORM-CD, a framework that combines a Variational Autoencoder (VAE) with a triplet loss, customized for change detection. The triplet loss enhances the accuracy of the approach while maintaining the computational and storage efficiency of VAE, making it particularly suitable for deployment on resource-constrained onboard hardware. We also introduce a new dataset -- STTORM-CD-Floods, annotated using a custom strategy tailored for change detection. Combining them resulted in significant performance improvements, as our method outperforms existing solutions in flood detection by significant margins, while the ability to detect other types of disasters was not significantly affected. Additionally, we highlight the potential of machine learning-free approaches and introduce new evaluation metrics to address testing challenges. These advancements bring us significantly closer to deploying a universal and accurate real-time disaster detection system in operational settings.