BRSMamba: Boundary-Aware Mamba for Forest and Shrub Segmentation From Diverse Satellite Imagery

Vegetation, as a critical ecological feature and irreplaceable material resource of the Earth's surface, plays a crucial role in environmental protection, resource assessment and urban planning. The rapid advancement of remote sensing platforms and sensor technologies has facilitated the acquisition of high-resolution remote sensing imagery, providing excellent conditions for the detailed analysis of vegetation. However, vegetation segmentation in remote sensing imagery poses distinct challenges, including extreme scale variance, spectral ambiguity and complex boundary characteristics. The performance of Convolutional Neural Network (CNN)-based methods in vegetation segmentation is constrained by their limited ability to model long-range spatial dependencies. In contrast, although Transformer-based architectures effective at capturing global context, their quadratic complexity makes them impractical for processing high-resolution remote sensing images. Recently, State-Space Models (SSMs) have emerged as a promising alternative owing to their linear complexity and efficient long-range modeling capabilities. Nevertheless, existing vision SSMs have two critical limitations: (1) insufficient modeling of boundary information; and (2) the limited performance improvements offered by multi-directional scanning strategies while increasing computational cost. To address these limitations, we propose BRSMamba, a boundary-aware network that integrates two novel modules with Non-Causal State-Space Duality (NC-SSD) to enhance both boundary preservation and global context modeling. Specifically, the Boundary-Subject Fusion Perception (BFP) module extracts robust boundary features via a Laplacian-of-Gaussian Convolutional Block (LCB) and fuses them with category-centric semantics to generate a boundary-subject map. This map then directs the Boundary-Body Resolution (BBR) module to inject boundary awareness into the NC-SSD state-transition matrix, enabling selective scanning that preserves fine details and captures long-range dependencies. Experiments on four benchmark datasets demonstrated that BRSMamba achieved 78.86% mIoU on Vaihingen, 82.84% on Potsdam, 54.37% on LoveDA, and 72.24% on GID while using only 10.61 M parameters and 26.28 G FLOPs. These results validate the effectiveness of the proposed method in balancing the accuracy and efficiency of complex vegetation segmentation tasks.