Large-scale flood monitoring based on time-series Sentinel-1 images and Z -score

Synthetic Aperture Radar (SAR) satellites have emerged as a crucial technology for real-time flood monitoring in the face of escalating flood disaster hazards. However, existing flood extraction techniques still face challenges in simultaneously achieving accurate urban inundation detection and extensive large-scale monitoring. To address this challenge, this study proposes a time-series Sentinel-1 SAR anomaly-detection approach based on an analysis of backscatter characteristics across different land-cover types and the integration of auxiliary datasets. First, an analysis of scattering behaviour in time-series SAR images revealed that, in natural areas, the consistency of backscatter intensity is strongly influenced by vegetation growth status. In building areas, rainfall and floodwater can intensify or suppress double-bounce scattering depending on building geometry and water depth, thereby disrupting intensity consistency. Based on these findings, a Z -score-based flood classification tree was developed. The method uses reference images and flood-period images to compute standardized anomaly maps with statistical interpretability under approximate normality and establishes operationally transparent flood-detection thresholds. The integration of VV and VH polarizations within the classification tree further improves the reliability of flood identification. Applied to the 2021 Weihui flood event, the method demonstrated strong performance, achieving a critical success index (CSI) of 60 % and overall accuracy (OA) of 90 % in natural areas, and a CSI of 62 % and OA of 73 % in building areas. The proposed approach shows significant advantages in accurately classifying flood-affected areas and offers the capability to monitor both large-scale floods and urban inundation.