Abstract/Summary

Cloud and snow often share comparable visual and structural patterns in satellite observations, making their accurate discrimination and segmentation particularly challenging. To overcome this, we design an innovative Transformer-guided architecture with complementary feature-extraction capabilities. The encoder adopts a dual-path structure, integrating a Transformer Encoder Module (TEM) for capturing long-range semantic dependencies and a ResNet18-based convolutional branch for detailed spatial representation. A Feature-Enhancement Module (FEM) is introduced to promote bidirectional interaction and adaptive feature integration between the two pathways. To improve delineation of object boundaries, especially in visually complex areas, we embed a Deep Feature-Extraction Module (DFEM) at the deepest layer of the convolutional stream. This component refines channel-level information to highlight critical features and enhance edge clarity. Additionally, to address noise from intricate backgrounds and ambiguous cloud-snow transitions, we incorporate both a Transformer Fusion Module (TFM) and a Strip Pooling Auxiliary Module (SPAM) in the decoding phase. These modules collaboratively enhance structural recovery and improve robustness in segmentation. Extensive experiments on the CSWV and SPARCS datasets show that our method consistently outperforms state-of-the-art baselines, demonstrating its strong effectiveness and applicability in real-world cloud and snow-detection scenarios.