Weakly supervised road network extraction for remote sensing image based scribble annotation and adversarial learning

Abstract

Performing accurate road extraction from remote sensing images has great practical importance. Nevertheless, the accuracy of segmentation can be seriously affected by complex and variable objective factors. Complex topological structures lead to discontinuity of extracted roads. State-of-the-art road extraction methods rely heavily on intensive annotation with high acquisition costs. In this work, we propose a scribble-based weakly supervised remote sensing road extraction network (WR2E), which can extract roads from remote sensing images based on scribble annotations. A road trimap generation algorithm is introduced to allow the propagation of semantic information for scribble annotations. The WR2E network contains two key modules, i.e., the road positioning module (RPM) and the oriented attention module (OAM). The road positioning module is designed to locate the initial position of the road from a global perspective, and the oriented attention module guides the propagation of high-level semantic information based on image affinity features. The proposed modal achieved the highest structural measure of 89.1%, the highest adaptive E measure of 92.0%, the highest weighted F measure of 88.6%, the highest mean absolute error of 3.6%, the highest IoU of 86.9%, and the highest F1-score of 93.7%, indicating the WR2E can accurately and effectively implement remote sensing image road extraction.