Road extraction from satellite images with iterative cross-task feature enhancement

Abstract

Recent study has shown that road orientation information is highly beneficial to the road extraction task. However, most existing approaches that integrate road extraction with orientation prediction in a naïve multi-task learning framework fail to explore their correlations, leading to unsatisfactory results with severe road discontinuity issues. In order to take full advantages of road orientation information to facilitate accurate road extraction, we propose an iterative cross-task feature enhancement network, which jointly optimizes the deep features for both road extraction and orientation prediction tasks in a mutually beneficial manner. To this end, we present two unique designs, namely, the semantic guided feature enhancement (SGFE) module and the orientation-aware feature aggregation (OAFA) module. The SGFE module is used to refine orientation prediction features by incorporating the semantic information of road extraction feature. Meanwhile, the OAFA module is able to enhance road extraction feature by adaptively deforming its receptive field according to the predicted road orientations. Our proposed network recursively alternates between these modules, permitting the extracted features to be collaboratively improved through cross-task information flow. We set new state of the art on two widely adopted benchmarks (67.00 road IoU on DeepGlobe [5] and 80.06 mIoU on Massachusetts [18]), which verifies the superiority of our method.11Our code is publicly available at:https://github.com/YinWeiling/Road-Extraction.