Feature-Separated Lightweight Model for Road Extraction in Wild Environments

Extracting road information from high-resolution remote sensing images is a critical challenge, particularly in complex environments with vegetation occlusion, shadow interference, and similar terrain textures causing discontinuous or incomplete road extraction. Existing methods often suffer from high computational costs and limited accuracy. To address these issues, we propose a lightweight road extraction model based on an encoder-decoder architecture, employing CSPDarknet53 as the backbone to reduce redundant computations. A Feature Separation and Enhancement Module (FSEM) independently processes spatial and channel features, while a Feature Aggregation Module (FAM) integrates multi-scale features for improved road continuity and accuracy. This design enhances prediction performance without significantly increasing parameters, reducing computational demands and making the model more adaptable for large-scale remote sensing tasks. Experimental results demonstrate that the proposed model achieves an mIoU improvement of 1.60% and 3.15% on the DeepGlobe and Mountain Road datasets, respectively, along with over 40% runtime reduction compared to other approaches. The proposed model combines high accuracy and efficiency, offering a scalable solution for complex road extraction tasks.