Dual-branch D-WNet: a classification method for cloudy remote sensing images integrating optical and time-series sar features

Multimodal fusion of optical imagery and time-series Synthetic Aperature Radar (SAR) data is essential for Land Use and Land Cover (LULC) classification in persistently cloudy regions. However, inherent spatiotemporal heterogeneity between modalities frequently impedes effective feature alignment, rendering conventional fusion techniques suboptimal in dynamic scenarios. To circumvent these constraints, we propose the Dynamic-Weighted Network (D-WNet), a dual-branch framework designed to explicitly decouple and adaptively synergize heterogeneous multimodal inputs. The architecture utilizes depthwise separable convolutions for the multiscale spatial–spectral analysis of optical imagery and ConvLSTM units to capture temporal scattering dynamics from SAR sequences. Furthermore, the proposed enhanced adaptive fusion module leverages a local-global dual-attention mechanism to dynamically integrate features, while a cross-layer residual guidance strategy refines boundary delineation and mitigates SAR speckle noise. Evaluations across diverse landscapes—including the Qinghai-Tibet Plateau, northwestern China, and California—demonstrate that the proposed framework outperforms existing methods, achieving overall accuracy improvements of 6.8-12.4%. By effectively fusing optical spectral features with time-series SAR responses, the D-WNet provides a robust solution for precise land-cover mapping in persistently cloudy environments.