A Comprehensive Satellite Review of Deep Learning Methods for Road Extraction from Satellite Images

Abstract

Given the increase in demand for the use of satellite imagery in automated and accurate extraction of roads, high-definition satellite imagery has recently created an increase in the need for automated and accurate techniques for extracting road networks from remote sensing images using modern techniques. Due to the increase in the accuracy and capability of creating strong representations within a deep learning architecture that can automate many of the manual tasks of creating and maintaining maps, deep learning has become the most prevalent method to extract roads from remote satellite images. In this paper, we have provided a complete overview of methods that utilize deep learning architectures to extract roads from satellite images. The papers that we discussed in this overview are organized by the architecture of the network or models, such as CNN’s, U-Net models, attention/implementation models, and transformer models. Performance comparisons for the extracted road data were performed based on the benchmarking datasets of DeepGlobe, SpaceNet, and the Massachusetts Roads Dataset using established performance metrics of Intersection Over Union (IoU), Precision, Recall, and F1 Score. Moreover, we discuss the strengths and weaknesses of the current techniques and methods used for extracting road data from these four types of models. We highlight common issues with extraction, such as road discontinuities, occlusion of roads, low contrast surfaces, and a lack of generalizability across datasets when using the former three types of networks, which cause similar performance issues. Finally, we propose several future research directions to develop better-performing systems that do not just rely on using deep learning with the aforementioned architectures but also on hybrid architectures, fusing multiple sensor data types and limiting the need to rely on collecting large sets of manually annotated satellite images in order to improve the performance of road extraction systems.