DeSPAR: Depth-Guided Semantic-Prompted Adaptive Refinement for ORSI Salient Object Detection

Optical remote sensing image salient object detection (ORSI-SOD) currently faces two major challenges: (1) existing RGB-based methods primarily depend on color and texture cues, which makes it difficult to obtain robust representations of object spatial structure under extreme imaging conditions; (2) significant morphological variations across object categories lead to semantic feature confusion in existing methods. To address these issues, we propose Depth-Guided Semantic-Prompted Adaptive Refinement (DeSPAR), a progressive refinement framework with geometric–semantic decoupling. To avoid excessive coupling between geometric and semantic signals in an end-to-end architecture, which would cause semantic priors to interfere too early with the construction of generic geometric representations, DeSPAR adopts a two-stage design for feature learning. In Stage 1, Depth-Guided Geometric Learning (DGL) employs a novel lightweight Depth-Guided Refiner (DGR) to build a generic geometric foundation. DGR utilizes RGB features to guide pseudo-depth denoising and injects geometric cues from pseudo-depth to enhance spatial feature representations. In Stage 2, Depth-Guided Semantic-Adaptive Refinement (DSR) inherits the encoder weights from DGL and introduces category-specific constraints. Under the guidance of a Semantic Prompt Bank constructed from DGL, DSR adaptively optimizes the representations of different categories through a prompt-guided mechanism. Experimental results demonstrate that DeSPAR surpasses 22 state-of-the-art methods on three public ORSI-SOD benchmarks, achieving superior performance with only 26.4M parameters while attaining an inference speed of 161 FPS.