Answers are grounded in this paper and cite their sources.
Land cover classification is a critical task in remote sensing image processing, with its core objective being the accurate classification of surface coverage through pixel-level feature extraction, thereby supporting applications such as environmental monitoring and land use planning. However, the inherent tension between the continuous physical nature of land cover and the discrete classification framework introduces semantic ambiguity at category boundaries and impedes effective model learning across scales. While edge features are commonly incorporated to mitigate boundary ambiguity, existing methods often ignore achieving synergistic integration between edge guidance and multi-scale feature fusion, leading to suboptimal performance. To address these challenges, we propose an Edge-Semantics Collaborative Network (ESCNet). We design a dual-branch architecture to decouple feature extraction across scales, and introduce an edge-guided fusion mechanism that strategically leverages edge cues to direct the integration of features from the dual branches, thereby establishing synergistic collaboration between boundary localization and semantic representation. First, we design a Direction-Sensitive Shuffle Detail Branch (DSSDB) to enhance spatial detail representation capabilities through multidirectional convolutions and channel rearrangement. Concurrently, a Spatial-Channel Dominated Semantic Branch (SCDSB) is constructed to explore synergistic spatial-channel contextual information using linear attention mechanisms. Additionally, an Edge-Guided Feature Fusion Decoder (EGFFD) is introduced to reduce fusion deviations between semantic and detail features by leveraging edge features as constraints. Experimental results on the DeepGlobe and LoveDA datasets demonstrate that ESCNet achieves mIoU scores of 73.9% and 59.32%, respectively, showing certain advantages in the preservation of fine-grained structures and semantic consistency.
No takes yet. Share an insight, caveat, or question.
Xie et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: