Machine learning–based flood inundation mapping using fused optical and SAR remote sensing: a case study of the Miyun flood

Against the backdrop of global warming, an intensified hydrological cycle has led to more frequent and more destructive short-duration extreme rainfall events. Rapid and objective delineation of flood inundation extent is therefore critical for emergency response and post-disaster assessment. Taking the extreme rainfall–triggered flood in Miyun District, Beijing, in July 2025 as a case study, this paper develops a machine-learning-based flood inundation mapping workflow through the fusion of optical and SAR remote sensing. Sentinel-1 and Sentinel-2 imagery acquired within 1 month before and 1 month after the flood event were collected, and pixel-wise monthly mean compositing was applied to reduce SAR speckle, cloud contamination, and scene-specific anomalies. Based on the composites, multi-source features—including SAR scattering characteristics, optical spectral information, and DEM-based terrain constraints—were integrated to perform supervised water/non-water classification, and newly inundated areas were extracted via pre-/post-event change detection. In addition, stable water (pixels classified as water in both periods) was identified and removed to avoid counting permanent water bodies such as the Miyun Reservoir as flood extent. High-resolution Jilin-1 optical imagery (1 m) was further incorporated for fine-scale boundary interpretation and auxiliary sample labeling. Results indicate that both SAR and optical data effectively capture flood-driven surface-water expansion, with newly inundated areas mainly distributed in low-lying zones around the reservoir and along both sides of river channels, forming a spatial pattern combining river-aligned belts and locally patchy clusters. Under the monthly compositing scheme, the flood inundation area estimated from Sentinel-2 is 1425 ha, while that from Sentinel-1 is 1284 ha; the optical estimate is 141 ha larger, corresponding to approximately +11.0% (with SAR as the reference). The difference reflects contrasting sensor responses under cloudy/rainy conditions, shallow inundation, and complex scattering environments, as well as the smoothing effect of monthly averaging on peak-flood signals. The proposed multi-source fusion and change-detection workflow provides a methodological reference for rapid flood mapping, damage verification, and flood risk management in Miyun and similar mountainous–river-valley composite regions.