Segment Anything Model Driven Cross-Hierarchical Fusion Network for Remote Sensing Images Semantic Segmentation

Semantic segmentation of remote sensing images (RSIs) involves dense, pixel-wise classification of high-resolution satellite images, serving as a foundational technique for applications including land cover mapping, urban analysis, and environmental monitoring. However, accurately delineating complex and diverse objects in RSIs is challenging due to the large domain gap from natural images and the heterogeneous characteristics of ground objects. In this article, we present a novel segment anything model driven cross-hierarchical fusion network (CHFNet) for remote sensing semantic segmentation. Specifically, a segment anything model-based encoder is utilized to extract comprehensive semantic features, which are further aligned with convolutional features through a cross-modal feature alignment module. To enhance semantic consistency and multiscale representation, a feature interaction fusion module is introduced for deep interaction and fusion of hierarchical features. Furthermore, a hypergraph-based decoder is designed to capture complex topological structures and inherent global relationships in RSIs. Extensive experiments on benchmark datasets (ISPRS Vaihingen and ISPRS Potsdam) demonstrate that CHFNet consistently outperforms state-of-the-art methods, and ablation studies further verify the effectiveness of each core component.