SegFormer-based segmentation approach for landscape planning and overhead remote sensing image analysis

This study developed a semantic segmentation framework for landscape element analysis. The objective was to achieve high precision, structural stability, and computational efficiency in spatial information extraction. The framework was built on a Transformer architecture and adopted SegFormer as the core modeling paradigm. A Mix Transformer (MiT) encoder generated multi-scale spatial feature representations, while a Multi-Layer Perceptron (MLP) decoder performed cross-layer semantic fusion. At the architectural level, an Edge-aware Auxiliary Supervision mechanism was introduced to strengthen structural learning. Gradient information served as an explicit constraint, improving discrimination of boundary-sensitive elements such as roads and water bodies. At the training level, a Unified Label Mapping strategy and a Boundary Jitter Augmentation strategy were implemented. These strategies enhanced generalization stability across multi-source landscape datasets. To eliminate performance bias arising from distributional differences and inconsistent training settings, a Two-track Fair Comparison Protocol was designed. This protocol enforced unified training conditions while maintaining independent evaluation. Experiments were conducted on the Land-Cover in High-Resolution Visual Dataset (LoveDA) and a self-constructed landscape dataset. Under the unified LoveDA setting, the model achieved a Mean Intersection over Union (mIoU) of 76.5% across five key landscape categories. This result represents a 5.3% improvement over DeepLabV3+. For the Boundary F1-score, structurally sensitive categories showed clear advantages, confirming the effectiveness of structural supervision. During the inference stage, the edge auxiliary branch used in training was removed. Under a unified setting of a 512 × 512 input and a batch size of 1, the deployed model achieved 42.1 Frames Per Second (FPS), indicating that the structural gains were not obtained at the cost of a significant loss in inference efficiency. Overall, the performance improvements were primarily driven by the interaction between the structure-adaptive workflow and planning-oriented evaluation metrics. Network scaling alone did not account for the observed gains. The proposed framework offers a reproducible and engineering-feasible solution for structured spatial representation in landscape planning, remote sensing interpretation, and geospatial information analysis.