Answers are grounded in this paper and cite their sources.
Land cover classification is essential for sustainable development but remains challenging due to the complexity of high-resolution remote sensing imagery, including spectral similarity between objects, variability within objects, and uneven spatial distributions. Existing methods often struggle to capture small-scale object details and maintain the integrity of large-scale structures. Additionally, the generalizability is limited, as many methods rely on training and validation within the same dataset. To address these challenges, we propose a novel multi-scale spatial-frequency learning framework (MSF-LCC), which introduces two key innovations: a spatial-frequency feature fusion module that effectively integrates spatial and frequency-domain features across multiple scales, and a selective spiking convolution to enhance decomposed low-, mid-, and high-frequency feature representation, respectively, while reducing computational redundancy. These contributions enable the model to better distinguish complex object boundaries and contextual relationships in heterogeneous scenes. We evaluate MSF-LCC on two independent test datasets, distinct from the training data, and demonstrate that it consistently outperforms six state-of-the-art semantic segmentation models. Specifically, MSF-LCC achieves a 12.41% improvement in mean intersection over union (mIoU) on the Guilin UAV dataset we constructed and a 6.06% improvement on the publicly available LoveDA dataset. These results highlight the robustness, generalization capability, and practical applicability of MSF-LCC for high-resolution land cover classification in real-world settings.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: