Edge-Guided and Multiscale Feature Optimization-Based Semantic Segmentation Network for Remote Sensing Images

Accurate semantic segmentation of remote sensing imagery is fundamental to geospatial analysis, yet the task remains daunting due to intrinsic complexities such as blurred object boundaries, severe occlusion, and high inter-class similarity. To overcome these persistent bottlenecks, this study presents an engineering-integrated framework that strategically synthesizes advanced vision components to balance segmentation accuracy with model efficiency. Rather than relying on a single core innovation, we rationally combine a Pyramid Vision Transformer (PVT) backbone with specialized edge and alignment modules to reconcile the trade-off between global context and local detail. Specifically, the architecture features an edge-guided encoder that augments the PVT backbone with a learnable Edge-guided Feature Module (EFM), explicitly targeting the recovery of boundary information often lost during downsampling. Complementing this, the decoder employs a Feature-aligned Pyramid Network (FaPN) reinforced by Adaptive Spatial Feature Fusion (ASFF). This combination ensures precise cross-scale alignment and dynamic feature selection, effectively suppressing background noise while preserving fine-grained spatial structures. Experimental results on two large-scale high-resolution remote sensing datasets demonstrate that the proposed network offers superior segmentation performance while maintaining an efficient balance between computational complexity and model parameters.