Multi-type road extraction and analysis of high-resolution images with D-LinkNet50

Road data form remote sensing is important for GIS modeling, vector analysis, and smart cities. Recently, there has been many scholars have successively combined deep learning with road extraction to meet practical needs. Based on the former research, this paper uses D-LinkNet50 which combines the pretrained LinkNet architecture with the dilation convolution. Training on the data set provided by DigitalGlobe, the results have shown that this D-LinkNet50 has achieved 83.1%, 79.7%, 81.3% in accuracy, recall, and F1-score, respectively, which is higher than that of D-LinkNet34 network 0.7%, 1.4%, 1.0%. So the extraction accuracy is significantly improved.