Conditional Diffusion-Guided Visual State Space Model for Semantic Segmentation of Optical Remote Sensing Images

Remote sensing image semantic segmentation is a fundamental task in geospatial information analysis, aiming to assign meaningful semantic labels to every pixel in high resolution satellite imagery. Despite significant progress made by Deep Learning (DL)-based models, existing methods still face challenges in accurately modeling complex object structures, maintaining semantic consistency, and recovering fine-grained details, especially in the presence of blurred boundaries, scale variation, and spectral mixing. In this paper, we propose Mam baSeg, a novel semantic segmentation framework for remote sensing images that deeply integrates conditional diffusion models with visual state space modeling. Specifically, we design an Edge-Guided Mamba (EGMamba) module that incorporates directional and edge-aware mechanisms into the state space modeling process, enhancing the model's ability to reconstruct multi-directional semantic structures and maintain boundary coherence. To further improve detail recovery, we introduce a Difference-Aware Feature Fusion (DiffFusion) module, which explicitly captures structural discrepancies between encoder and decoder features, enabling precise edge localization and robust small object segmentation. In addition, a Frequency Feature Enhance (FreEnhance) module is incorporated in the diffusion decoding phase, leveraging Fourier domain modeling and cross-frequency attention to compensate for texture degradation and strengthen high-frequency detail expression. Extensive experiments on multiple benchmark datasets demonstrate that MambaSeg achieves superior performance in terms of structural integrity, small object recognition, boundary restoration, and se mantic consistency, surpassing state-of-the-art methods. The code is available at https://github.com/darkseid-arch/RS-MambaSeg.