Decoupling Representation Learning for Remote Sensing Semantic Segmentation

Semantic segmentation plays a pivotal role in the interpretation of remote sensing entities. However, the complexity of the environments, diversity of objects, and richness of details make semantic segmentation even more challenging. Existing methods have limitations in interclass edge continuity and intraclass completeness within segmentation results. To address these issues, a feature decoupling guided Network is developed for learning the discriminative representation. In which the feature decoupling module separates encoded features into homogeneous information for primary object features and distinct information for object boundaries. Another, the semantic-aware integration unit is employed to strengthen semantic consistency during decomposition. To facilitate practical application, we created the Taiyuan Land Cover (TYLC) dataset for semantic segmentation to analyze land resource utilization. Extensive experiments on the TYLC dataset achieved a mean intersection over the union of 54.2%, the mean $F_{1}$ score of 67.7%, and the mean recall of 66.6%. Quantitative results demonstrate the algorithm’s superiority, and visualizations indicate that the segmentation output has excellent completeness and edge continuity.