Evaluation and comparison performance of deep neural networks FCN and RDRCNN in order to identify and extract urban road using images of Sentinel-2 with medium spatial resolution

Road extraction using remote sensing images has been one of the most interesting topics for researchers in recent years. Recently, the development of deep neural networks (DNNs) in the field of semantic segmentation has become one of the important methods of Road extraction. In the Meanwhile The majority of research in the field of road extraction using DNN in urban and non-urban areas has been done using images with high spatial resolution. In this research, for the first time, to extract the road using DNN, the images with medium spatial resolution of Sentinel-2 sensor were used, so that the image of Tehran as a test data and from 7 other cities (Mashhad, Isfahan, Shiraz, Tabriz, Kermanshah, Urmia and Baghdad) were used as training and validation data. In the Meanwhile, after preparing and labeling all the pixels related to the road surface, the images are converted into 256 × 256 pieces, and after separating the unsuitable parts, for test, training and validation data, respectively. 135, 1500 and 100 image pieces were obtained. Finally, deep refined residual convolution neural networks (RDRCNN) and U-Net, which are based on fully convolutional networks (FCN), were used to train and extract the road complication. The results show that both RDRCNN and FCN models have well identified and extracted Tehran urban road network from Sentinel 2 images in comparison with the ground reality data. Meanwhile, the FCN model performed better than the RDRCNN model both visually and in terms of accuracy assessment metrics, so that for the FCN model, the criteria Recall 82.92%, accuracy 77.67%, F1 score 77.53 and overall accuracy 96. 30% and for RDRCNN the criteria Recall were 80.43%, accuracy 71.37, F1 score 72.14% and overall accuracy 95.71%. In general, the findings of this study show the potential of using DNN methods to extract urban roads using images with medium spatial resolution of Sentinel-2.