Abstract: Abstract: Semantic segmentation of remote sensing images is an important task in remote sensing image processing. However, significant variations in object scales and complex spatial structures in remote sensing scenes limit the ability of existing single feature extraction approaches to adequately capture both local structural details and global contextual information. Although existing CNN–Transformer hybrid models can capture local structures and global semantic information simultaneously, insufficient interaction between heterogeneous features limits effective feature fusion. To address these issues, a multi-view dual-stream interactive fusion network is proposed for remote sensing image semantic segmentation. First, ResNet-50 and SwinTransformer are employed in parallel to extract local and global features, respectively. Second, grouped convolution and pointwise convolution are introduced to facilitate feature interaction between the two branches, while grouped feature aggregation and global semantic perception are employed to enhance high-level semantic representations. Finally, multi-scale gated attention is used to optimize feature selection and suppress background noise. Experiments on three public datasets, including Potsdam, LoveDA, and UAVid, show that the proposed method achieves mIoU values of 87.3%, 53.3%, and 70.2%, respectively, demonstrating its effectiveness for remote sensing image semantic segmentation.