Road Detection with Deep Learning in Satellite Images

Road detection from high-resolution satellite images using deep learning is proposed in this article. The VGG19 architecture, which is one of the deep convolutional neural network architectures, is used in the proposed method. To detect the road, two steps are implemented. To achieve high accuracy, image segmentation is done in the first step. At this stage, based on the semantic division, the objects whose area is small are removed. In the second stage, edge detection of images combines two techniques of segmentation and edge detection to improve road detection. Considering the good accuracy of the VGG19 architecture and the need for few parameters, the obtained results are favorable. To check the performance of the proposed method, the IoU criterion was used. The values obtained for this criterion show an improvement of more than 80%. While this criterion is less than 80% for the compared methods. The obtained results can be used for the purposes of digital mapping, transportation management and many other applications.