Deep Learning Architecture U-Net Based Road Network Detection from Remote Sensing Images

Abstract

Roads are the foundation of human civilisation and one of the most important routes of transportation. For the city planning, vehicle traffic control, road network monitoring, map updating and GPS navigation, the study of road extraction is extremely important. Due to similar spectral characteristics, occlusion of buildings and trees present in remote sensing images makes to extract the road surface is challenging task. This paper address the road network detection based on deep learning sementic segmentation architecture such as U-Net and SegNet from Remote Sensing Images (RSI). Publically available dataset is used to train the U-Net and SegNet. These methods are tuned with various hyper parameters such as learning rate, batch size and epochs. The performance of the methods is also observed under various optimization algorithm like SGD and ADAM. The suggested method performance is measured by training and testing accuracy, total training time, inference time, average iou score and average dice score.