SFMRNet: Specific Feature Fusion and Multibranch Feature Refinement Network for Land Use Classification

Abstract

Land use classification of high-precision satellite images using semantic segmentation methods has become mainstream. In this field, global context information plays an irreplaceable role. However, most current methods struggle to effectively utilize this global context, which results in low segmentation accuracy, especially in scenes with similar objects, small targets, or obscured by shadows. To address the above issues, this article introduces SFMRNet—the network that integrates the advantages of Transformer and convolutional neural network (CNN)—and designs various modules to utilize the power of contextual information as much as possible. First, we design a specific enhanced feature fusion module (SEFFM) that selectively enhances spatial or channel information of feature maps before fusion, effectively mitigating small interclass differences. Second, our proposed multibranch feature refinement module (MFRM) facilitates the interaction between different feature layers and refines these features to enhance multiscale characterization. This improves the segmentation of small-sized targets and addresses the occlusion issues. Finally, comprehensive testing and detailed ablation analysis are conducted on three datasets: the ISPRS Vaihingen, ISPRS Potsdam, and LoveDA land use classification datasets. The results demonstrate that SFMRNet exhibits superior segmentation capabilities compared to existing advanced methods.