SSPNet: Spatial-Semantic Foundation Prior Network for Remote Sensing Few-Shot Segmentation
Abstract
Few-shot segmentation seeks to identify and delineate target objects in a query image with the guidance of only a handful of annotated support samples. While considerable progress has been made on natural image benchmarks, applying these techniques to remote sensing imagery remains challenging due to the pronounced intraclass appearance diversity and the pervasive spectral confusion between foreground and background. In this article, we propose a spatial-semantic foundation prior network (SSPNet) that harnesses two heterogeneous frozen foundation models to address these difficulties. SSPNet comprises a dual-stream multimodal semantic fusion (DMSF) module and a dynamic class activation map (CAM)-guided mask-aware cross-attention (DCMA) module. The DMSF module employs a frozen DINOv2 encoder and a frozen GeoRSCLIP text encoder whose representations originate from different pretraining objectives; a learnable cross-modal projection maps the language representation into the spatial feature space, enabling complementary spatial and semantic CAMs to be constructed in a shared hidden space. The DCMA module then incorporates the fused activation prior through bounded residual modulation and foreground-constrained cross-attention. An auxiliary pseudomask objective provides a downstream-to-prior gradient pathway through the learnable activation-generation and feature-refinement layers, coupling their optimization during episodic training. Experiments on the iSAID-5$^{i}$ and LoveDA-2$^{i}$ benchmarks show that SSPNet achieves state-of-the-art or competitive performance across the evaluated settings, with the most pronounced improvements under the one-shot protocol.