Unsupervised Continuous Flood Mapping From Multitemporal, Multiview, and Multisensor SAR Sequences

Floods are among the most disruptive natural hazards, requiring emergency managers and civil protection authorities to access timely and reliable information on flood extent and its evolution over time. Satellite-based flood mapping plays a critical role in supporting emergency response and preparedness, particularly when ground access is limited. Synthetic aperture radar (SAR) data are well suited for this purpose, as they provide day-and-night observations independent of cloud cover. However, continuous flood monitoring remains challenging due to irregular revisit times, limited availability of suitable preflood reference images, and the need to combine data from different sensors while ensuring fast product delivery to end users. To address these challenges, this article presents an automated and sensor-flexible framework for continuous flood monitoring using multitemporal SAR data. The methodology implements a unified workflow that exploits postevent feature-based mapping and change-detection analysis, allowing flood extent estimates when suitable preflood reference images are available as well as when such references are missing or inconsistent. The workflow operates directly on heterogeneous SAR acquisitions (e.g., Sentinel-1 GRD and COSMO-SkyMed Level-1D products) and is designed to remain robust under varying imaging geometries and operational constraints. To this aim, flood detection is performed through unsupervised clustering of SAR backscatter features, with optical data used, without human intervention, to semantically guide the cluster interpretation. The methodology was tested on two contrasting flood scenarios characterized by different temporal dynamics. The May 2023 Emilia–Romagna flood represents a fast-onset event, with daily flood maps generated throughout the emergency period. The La Mojana floodplain in Colombia represents a markedly different setting, characterized by long-lasting and recurrent inundations. Validation against benchmarks shows high and consistent performance of the proposed approach, with overall accuracy close to 98% and F1-scores ranging from approximately 84% to over 91% across different flood scenarios and sensors. Overall, these results demonstrate that the proposed framework enables reliable and temporally consistent flood monitoring across sensors, time scales, and flood regimes.