Thin Cloud Removal From Remote Sensing Imagery Using Difference Map Supervision and Band Balancing

Thin cloud contamination constitutes a formidable barrier to the effective application of optical remote sensing imagery. Pixels that are affected by such contamination retain partial information, with thin clouds exerting differential effects across spectral bands. Motivated by these observations, we propose DBM, a lightweight external module for thin-cloud removal. DBM introduces a temporally adjacent auxiliary image into the MPRNet backbone and regulates auxiliary-information fusion through difference map supervision and band balancing, resulting in DBM-MPRNet. DBM incorporates a difference map supervision module (DMSM) as the primary design and a band balancing module (BBM) as the complementary design. DMSM generates the difference map and transforms it into a learnable spatial gate to regulate auxiliary information along the spatial dimension. BBM dynamically adjusts the weights of the spectral bands during the training process to regulate auxiliary information along the spectral dimension. Through this cooperative design, DBM exploits reliable information from the auxiliary image while suppressing interference from erroneous information caused by temporal inconsistency. Extensive experiments on simulated and real-scene datasets demonstrate that DBM-MPRNet achieves significantly better quantitative and visual results than the baseline methods while maintaining a low computational cost. In addition, DBM shows strong robustness and maintains favorable cloud removal performance even when the auxiliary image is severely corrupted by noise. As a lightweight external module, DBM also exhibits strong transferability. When integrated into MemoryNet or MSDA, it significantly improves the cloud removal capability of the corresponding framework. These results demonstrate that DBM provides an effective and practical solution for thin-cloud removal in remote sensing imagery.