DBRSNet: a dual-branch remote sensing image segmentation model based on feature interaction and multi-scale feature fusion
Yong Ji, Wenbin Shi, Jingsheng Lei, Jiayin Ding
Abstract
High-resolution images encapsulate abundant geographical information; however, precise semantic segmentation is essential for effective remote sensing image interpretation. Remote sensing semantic segmentation categorizes pixel-level image information into distinct land cover types, providing essential support for urban planning, resource management, and environmental monitoring. However, existing approaches encounter two major challenges: insufficient retention of fine-grained local details and suboptimal global contextual modeling, especially in intricate and high-resolution remote sensing scenarios. These limitations result in fragmented object boundaries, degradation of small-scale structures, and challenges in comprehending large-scale spatial dependencies. To address these limitations, we introduce DBRSNet, an advanced dual-branch remote sensing segmentation framework that integrates feature interaction with multi-scale feature fusion. In DBRSNet, the Feature-Guided Selection Module (FGSM) adaptively integrates complementary features from CNN and Transformer branches, while the Convolutional Attention Integration Module (CAIM) enhances global dependencies and spectral correlations, ensuring a more comprehensive feature representation. Extensive evaluations on the ISPRS Vaihingen and ISPRS Potsdam datasets validate that DBRSNet surpasses 14 cutting-edge remote sensing segmentation models across all assessment metrics, highlighting its exceptional performance and competitiveness.