Flood monitoring: An innovative application of multisource image fusion and transfer learning

The intensification of global climate change and the increasing frequency of extreme weather events require advanced methods for flood disaster monitoring and management. We propose a cross-sensor framework for robust water body mapping using multitemporal optical imagery, mitigating radiometric discrepancies and temporal drift. By integrating satellite datasets from multiple platforms (Sentinel-2A, Landsat-8, Gaofen-1 (GF-1), and Huanjing-1A (HJ-1A)), we constructed a harmonized remote sensing time series database, developed a relative radiometric normalization approach for heterogeneous imagery, and designed a spatiotemporally invariant feature extraction mechanism. Key innovations in this study include the implementation of a particle swarm optimization-enhanced random forest (PSO-RF) classifier to enhance model robustness and the development of cross-sensor sample transfer strategies to improve the interoperability of training data. To validate the proposed framework, we performed water body extraction using a fusion of multidimensional features, incorporating spectral indices (e.g., the Normalized Difference Water Index (NDWI)) and spatial-textural attributes as reference benchmarks. Comparative validation against traditional NDWI-based methods revealed significant classification accuracy improvements, with Landsat-derived water body extractions exhibiting 2.03%-13.65% increased precision. The postnormalized Sentinel-2A and GF-1 imagery consistently yielded Kappa coefficients above 0.8, and the improvement in classification accuracy ranged from 1% to 2%. The proposed framework demonstrated significant operational efficiency and detection accuracy improvements for flood monitoring in arid regions. These advancements provide a solid technical foundation for dynamic water resource monitoring in the context of climate change, offering direct implications for real-time disaster management applications.