A Bio-inspired Geospatial Reasoning Framework for Referring Remote Sensing Image Segmentation

Abstract

Referring remote sensing image segmentation (RRSIS) aims to localize and segment specific geospatial targets in images guided by natural language expressions. Although existing methods have achieved promising progress in vision-language alignment, they still suffer from three major limitations in complex scenarios, including interference from densely distributed similar targets, insufficient spatial relations modeling, and difficulty in fine-grained segmentation of objects with weak boundaries. To address these issues, this paper proposes BioGeo-RIS, a novel bio-inspired dual-stage framework for RRSIS built on the paradigm of "cognitive anchoring–perceptual enhancement." First, a spatial cognitive alignment module is designed to emulate the ventral and dorsal streams of the human visual system. This module explicitly models attribute and spatial-relation semantics in referring expressions, and generates task-oriented prompts via multi-branch cross-modal interactions, effectively resolving reference ambiguity caused by complex expressions and similar targets. Second, a foveal perceptual enhancement module is proposed and embedded into the SAM2 encoder. By decoupling high- and low-frequency features and applying adaptive residual fusion, this module enhances global structural features and local boundary details, greatly improving mask prediction quality for weak-boundarytargets, small objects, and complex backgrounds. Experiments are conducted on three public datasets, namely RRSIS-D, RIS-LAD, and RRSIS-HR. The results show that BioGeo-RIS outperforms state-of-the-art methods across various metrics, particularly demonstrating clear advantages in high-threshold precision and mean intersection over union, which strongly validate the effectiveness of the proposed method for complex remote sensing referring segmentation tasks.