Road Extraction from High-Resolution Remote Sensing Images Based on EDRNet Model

Abstract

The existing methods for extracting the road parts from high-resolution remote sensing images are limited by the incomplete extraction results and poor boundary quality.To address the problem, a new method based on the EDRNet model is proposed for extracting road parts from remote sensing images.The residual network is used to build the road extraction model, EDR1, which retains the detailed information of the road and accelerates the network convergence.Then multi-scale and multi-level road feature information is fused to design a model, EDR2, for optimizing the road extraction results.On this basis, the mixed loss function is designed to make the road extraction results more complete.Experimental results on the Maschusetts Roads dataset show that the recall rate, precision and F1-score of the proposed methods reach 84.4%, 81.7%, and 82.9% respectively.The proposed method can provide complete and accurate extraction results.