Multiscale Long-Distance Feature Aggregation Network for Geospatial Semantic Segmentation in High-Resolution Remote Sensing Imagery

High-resolution remote sensing semantic segmentation is a fundamental task in Geospatial Artificial Intelligence (GeoAI). Existing CNN-based methods are effective for local and multiscale feature extraction but often lack progressive cross-scale semantic propagation, while attention- and Transformer-based methods improve global spatial modeling but generally ignore frequency-domain regularities. To address these limitations, this study proposes a Multiscale Long-Distance Feature Aggregation Network (MLFANet), a unified spatial-frequency segmentation framework for high-resolution remote sensing imagery. MLFANet introduces three key components: a Multiscale Global Dependency Extraction module for cascaded cross-scale contextual refinement, an FFT-based frequency-domain branch with learnable global filtering for capturing structural and texture regularities, and a bidirectional Spatial-Frequency Fusion module for adaptively aligning spatial details with frequency responses. Experiments on the ISPRS Potsdam and Vaihingen datasets demonstrate the effectiveness and feasibility of the proposed model. MLFANet achieves AF, MIoU, and OA values of 86.03%, 76.21%, and 88.70% on Potsdam, and 83.17%, 71.90%, and 86.33% on Vaihingen, respectively, outperforming representative CNN-based, attention-based, and hybrid models in overall metrics. In terms of computational complexity, MLFANet requires 17.49 G FLOPs under an input size of 256 × 256 pixels, indicating its practical feasibility for patch-based high-resolution remote sensing segmentation. Ablation studies further verify that multiscale dependency extraction, frequency-domain modeling, and adaptive spatial-frequency fusion each contribute to the final performance.