
Urban snow mapping in complex built environments faces three interconnected challenges: fragmented snow cover, variable SAR responses across urban surfaces, and deviations between reference snow boundaries and SAR structural boundaries. Existing snow products and datasets offer limited support for evaluating these urban-specific difficulties. To address this gap, we construct UrbanSnow-NE10, a 10 m multisource benchmark covering five cities in Northeast China. The dataset integrates Sentinel-1 SAR, Sentinel-2 optical imagery, and topographic data, and characterizes mapping complexity through three dimensions: spatial dispersion of snow cover, SAR heterogeneity within snow regions, and label-SAR boundary deviation. Building on this benchmark, we propose SAR-SnowNet, a multimodal segmentation network with three complementary components: a Cross-modal Spatial Modulation Block for adaptive optical-SAR feature interaction, a Semantic Spatial Refinement Block for urban texture suppression, and a SAR Confidence Boundary Supervision Module for reducing the influence of uncertain boundary labels. A spatially independent cross-city protocol is adopted, with Shenyang reserved for testing. SAR-SnowNet achieves an F1-score of 95.43%, an IoU of 91.26%, and a Trimap-IoU of 74.63%. Complexity-stratified evaluation further demonstrates improved robustness under increasingly challenging urban conditions. The dataset and framework provide a reproducible benchmark and a physically constrained boundary-learning paradigm for urban snow monitoring.
Remote sensing change detection provides an essential foundation for diverse practical applications, including urban development, post-disaster analysis, and ecological management. Existing change detection models still suffer from information degradation during upsampling and insufficient representation of edge details, which manifest as missed detections, blurred edges, and fragmented topological structures. To address the above issues, this study proposes a pixel-level change detection network that integrates optical flow-guided feature fusion with edge prior constraints. To correct spatial misalignment between cross-scale features, an optical flow‑guided Feature Fusion Module is constructed to predict dense sampling displacements from the joint representation of deep and shallow features. The spatial responses of the deep and shallow features are adaptively adjusted through differentiable resampling, while a dynamic fusion mechanism is employed to integrate information from different hierarchical levels. In this way, local spatial misalignment across hierarchical features is alleviated, thereby facilitating the effective fusion of deep semantic information and shallow structural details. Additionally, to enhance boundary perception, this paper constructs a multi-task learning-based edge detection branch. By incorporating edge priors into the network via the edge cross-attention mechanism, this branch effectively enhances the model’s ability to perceive complex geometric boundaries. Furthermore, a Global Feature Refinement Module combines multi-dimensional attention with channel shuffling to dynamically recalibrate feature responses, suppress background noise retained during optical flow-guided feature fusion, and enhance discriminative feature representations. Comprehensive evaluations across multiple benchmarks demonstrate that the proposed approach achieves competitive performance compared to representative baselines, particularly excelling in building change detection in terms of overall accuracy , F1 score, and boundary localization quality.