A road semantic segmentation system for remote sensing images based on deep learning

Abstract

With the rapid development of deep learning of computer science nowadays in China, many fields in academic research have experienced the powerful and efficient advantages of deep learning and have begun to integrate it with their own research. To be specific, in the field of remote sensing, the challenge of road extraction from the original images can be effectively solved by using deep learning technology. Getting a high precision in road extraction can not only help scientists to update their road map in time but also speed up the process of digitization of roads in big cities. However, until now, compared to manual road extraction, the accuracy is not high enough to meet the needs of high-precision road extraction for the deep learning model because the model cannot extract the roads exactly in complex situations such as villages. However, this study trained a new road extraction model based on UNet model by using only datasets from large cities and can get a pretty high precision in extraction for roads in big cities. Undoubtedly, this can lead to over-fitting, but its unique high accuracy ensures that the model's ability to extract roads can be well utilized under the situations of large cities, helping researchers to update road maps more conveniently and quickly in large cities.