An Improved Deep-Learning Model for Road Extraction from Very-High-Resolution Remote Sensing Images

Abstract

Road extraction from remote sensing images based on deep learning has always been a hot research topic. However, due to the complexity of road conditions, many deep-learning models cannot obtain satisfactory results of road extraction. To improve the accuracy of road extraction, this paper proposes an improved deep-learning model based on the structure of Deeplabv3+. The proposed model uses four blocks with ResNeSt and ASPP to extract the feature maps, which can improve the integrity of the extracted roads. And each of the extracted low-level features is transmitted to the decoder. The decoder focuses on the effective fusion of low-level and high-level feature maps and gradually restores this information layer by layer, which can produce a more accurate result of segmentation. And because the road results extracted by neural networks often have broken road lines, this paper also proposes a data post-processing method to effectively solve this problem. The final experimental results show that the proposed model has further improvements in accuracy, mean Intersection over Union, and Fl-score compared to some other state-of-the-art(SOTA) models.