Automatic Road Extraction from High-Resolution Remote Sensing Images Using a Method Based on Densely Connected Spatial Feature-Enhanced Pyramid

Abstract

Road extraction is an important task in remote sensing image information extraction. Recently, deep learning semantic segmentation has become an important method of road extraction. Due to the impact of the loss of multiscale spatial features, the results of road extraction still contain incomplete or fractured results. In this article, we proposed a deep learning model, which is called the dense-global-residual network that reduces the loss of spatial information and enhances context awareness. In the dense-global-residual network, the residual network is used to extract the features at different levels. To obtain more abundant multiscale features, a dense and global spatial pyramid pooling module based on Atrous Spatial Pyramid Pooling is built to perceive and aggregate the contextual information. The proposed method obtains better results on the GF-2 road dataset and public Massachusetts road dataset of aerial imagery. In order to prove the effectiveness of our method, we compared with four methods, such as DeepLabV3+, U-net, D-LinkNet, and coord-dense-global model, and found that the accuracy of our method is considerably better. Moreover, the dense-global-residual network can also effectively extract roads, especially trees and building shadows that occlude the road. In addition, our method can successfully extract roads in regions of different development levels in universality experiments. This indicates that the proposed method can effectively maintain the completeness and continuity of roads and improve the accuracy of road segmentation from high-resolution remote sensing images.

Keywords

• Deep learning
• dense and global spatial pyramid pooling (DGSPP) module
• remote sensing image
• road extraction