Abstract
Satellite imagery can play a crucial role in disaster management, but critical images often take hours or even days to reach end-users, and upgrading hardware to improve transmission speed is prohibitively expensive for many small satellite missions. This article thus explores onboard change detection methods as a cost-effective alternative to reduce reaction time. Building on RaVAEn, we introduce STTORM-CD, a framework that combines a Variational Autoencoder (VAE) with a triplet loss, specifically designed for change detection. The triplet loss improves detection accuracy while maintaining the computational and storage efficiency of VAE, making it suitable for deployment on resource-constrained satellite hardware. To support training and evaluation, we present a new dataset, STTORM-CD-Floods, annotated with a custom strategy optimized for flood detection, along with new metrics, AURC and RDP, designed to address limitations of RaVAEn evaluation strategies, which are influenced more by dataset composition than model performance. Our experiments show that STTORM-CD outperforms existing flood detection methods, achieving an increase of 35 percentage points (pp) in custom AURC and standard AUPRC metrics against RaVAEn on the presented STTORM-CD-Floods dataset, while showing negligible changes in AURC (approximately -4 to +0.1 pp) for landslides and wildfires. This demonstrates that improvements on one disaster type do not necessarily compromise performance on others and highlights the potential for a universal and accurate real-time disaster detection system.