A Road Extraction Network Focusing on Long Range Relationships and Directional Linear Features of Roads

Extracting road information from remote sensing images is an effective method widely applied across multiple domains. Roads in remote sensing images exhibit characteristics such as narrowness, extended spans, meandering patterns, and interconnected intersections. The comprehensive extraction of road information remains a significant challenge due to occlusion effects from elements like trees and buildings, coupled with the inherent complexity of the road background. In response to this challenge, we propose a novel road extraction network, referred to as LKDSNet. It combines large kernel convolutions with a dilated pyramid module, enabling the learning of long range relationships of roads and facilitating multiscale fusion and transmission. Directional stripe convolution modules are incorporated into skip connections to model and extract directional and linear features of roads. Integrated into the decoding stage, these modules compensate for feature losses and enhance feature representation. Extensive experiments on the DeepGlobe dataset validate that our approach successfully extracts more complete road information with optimal performance, achieving higher IoU of 0.732 and F1 score of 0.845.