Semantic Segmentation of Land Cover in Remote Sensing Images Based on Multiscale Feature Adaptive Enhancement

High-resolution remote sensing images (HRRSI) exhibit significant differences in the scale of objects, and effectively extracting multi-scale features is key to describing objects of different sizes. Existing methods mostly aggregate features from specific levels, making it difficult to balance local details and global dependencies. To address this, this paper proposes MFESegNet, a land cover semantic segmentation network based on an encoder-decoder architecture, aimed at synergistically enhancing multi-scale feature representation and adaptive context modelling. The encoder uses a dual-branch architecture. The Adaptive Grouping Gated Convolution Block (AGGCB) uses innovative hierarchical channel grouping and dynamic gating mechanisms to suppress background noise and extract fine-grained local details. The Dual Attention Efficient Transformer Block (DAETB) uses spatial-channel decoupled attention and a frequency-domain feedforward network to efficiently model long-range dependencies while preserving high-frequency edge information. The Global-Local Feature Fusion (GLFF) module incorporates bidirectional attention gating to bridge the semantic gap between local and global features, enabling adaptive weight calibration for the two branches. The Multi-Scale Feature Attention Fusion (MFAF) module resolves cross-level semantic misalignment through synergistic, semantic-aligned convolutions and dual attention, thereby enhancing the model's adaptability to multi-scale objects. The decoder uses Mamba with a deformable scanning strategy (DSSM), which dynamically adjusts the sequence scanning path of Mamba based on the target geometry, achieving high-precision reconstruction of irregular boundaries. We carry out extensive comparison and ablation experiments on three representative datasets. The results show that compared with other methods, the MFESegNet has significant advantages in multi-scale feature extraction and contextual information representation, and provides a more effective solution for semantic segmentation of land cover.