Multisensor Change Detection in Flood Monitoring: A Reliability-Aware Dual-Regularized Fusion Network

Abstract

Change detection (CD) from bitemporal remote sensing images is essential for flood disaster assessment. Recent advancements in Earth observation have enabled multisensor CD (MSCD) to exploit the complementarity of heterogeneous data sources, such as synthetic aperture radar (SAR) and multispectral (MS) images. However, flood-related MSCD remains challenging because flood events introduce reliability imbalance between SAR and MS observations, local feature inconsistencies, and the representation bottleneck encountered in multimodal learning. To tackle these challenges, this article proposes a dual-regularized SAR-guided fusion network. The framework leverages the structural stability of SAR features to adaptively refine MS feature representations, enabling reliability-aware multimodal fusion. To mitigate flood-induced local feature inconsistencies caused by dynamic surface variations, a motion-aware feature enhancement module is designed to enhance bitemporal feature representations. In addition, a dual self-supervised regularization strategy is introduced to balance multimodal semantic consistency with information preservation. Moreover, five representative MSCD architectural paradigms are standardized to support systematic evaluation and comparison. Experiments on a public multisensor dataset and four cross-regional flood scenarios demonstrate the effectiveness and generalizability of the proposed method.