Function-Decoupled Environment-Gated Multimodal Fusion for Chlorophyll-a Retrieval in a Cold-Arid Shallow Lake

Lake Ulansuhai is a eutrophic shallow lake in the cold-arid region of northern China, where shallow water depth, frequent wind disturbance, and heterogeneous optical conditions complicate satellite-based chlorophyll-a (Chl-a) retrieval. To improve retrieval under such complex observation conditions, this study develops a function-decoupled environment-gated fusion framework (FDEGF) that integrates optical, synthetic aperture radar (SAR), and meteorological observations through predefined computational roles. Sentinel-2 optical features form the direct retrieval pathway, Sentinel-1 SAR features provide auxiliary information, and ERA5 meteorological variables condition the participation of the auxiliary pathway at the sampling-event level. A satellite-ground matched dataset containing 287 valid observations from 2017 to 2019 was constructed for Lake Ulansuhai. Across 31 paired random train-test splits, FDEGF showed a generally positive absolute-error advantage relative to simplified fusion structures, reducing median MAE by 0.360-0.467μgL-1. Further analyses showed that model performance depended on the correspondence between event-level meteorological gating and the SAR auxiliary pathway, while the three input pathways exhibited differentiated model dependence. Under unseen-station validation, FDEGF retained positive spatial transferability within the lake. However, leave-one-sampling-event-out validation yielded a pooled R2 of -0.144 and no consistent advantage over the structural controls. These results indicate that explicitly separating direct retrieval, auxiliary information, and meteorological gating provides a structurally explicit strategy for multimodal Chl-a retrieval within the represented sampling-event domain, while transfer to previously unseen sampling events remains limited.