Analysis of Methods, Challenges, and Perspectives in Rural Road Networks Detection with a Focus on Remote Sensing Images

Background and Objective: By advancement of remote sensing and deep learning technologies, the automatic identification of road networks, particularly in rural areas and secondary roads, has become feasible. Moreover, traditional mapping methods, due to their high cost and time-consuming nature, have been increasingly replaced by approaches based on remote sensing data and machine learning. Therefore, this study aims to comprehensively review research conducted on the preparation of road maps using remote sensing data, especially in recent years. In general, preparing a map of the road network involves various methods, and one of the reliable and cost-effective methods is automatic road detection using remote sensing images. Examination of various research results indicates that among the available automatic approaches, methods based on deep learning networks can provide acceptable and more reliable accuracies compared to other conventional methods. Considering the common width used in road construction, remote sensing images with different spatial resolutions have been used in research, each with its advantages and disadvantages. In general, for the type of route under study, which is the rural road network, satellite images with medium spatial resolution and free access cannot achieve high accuracy. Thus, satellite images can be integrated to enhance spatial resolution and improve detection algorithms. One way to integrate satellite images is through super-resolution algorithms. This study can serve as a reference for comparing and selecting methods for the automatic detection of road networks and for improving the spatial resolution of remote sensing images to assist in identifying narrow roads (such as rural road networks), so that researchers can select appropriate data and algorithms based on the objective and type of road network under study.

Materials and Methods: This research aims to investigate the existing methods for road network detection and the utilization of satellite images with medium spatial resolution for this purpose. Initially, the data and methods applicable for generating a road network map were examined. Subsequently, the principles used in the field of road network detection using remote sensing images were described, and based on these principles, classification methods, segmentation, road index, and machine learning were implemented. Methods for improving the spatial resolution of satellite images were also investigated for employing satellite images with medium spatial resolution. Finally, the methods were reviewed in terms of input parameters, mechanism, and output, to identify their strengths and weaknesses and to utilize them optimally for various applications.

Discussion and Analysis: According to the reviews of the examined articles from various Authentic journals in the field of road network detection, classification, segmentation, road index, and machine learning methods account for approximately 28%, 31%, 5%, and 36% shares, respectively. In recent years, classification and segmentation methods based on neural networks have been developed, encompassing a larger share (about 60%) of machine learning methods in general. Additionally, in super-resolution, investigations show that methods based on traditional techniques and deep learning account for approximately 44% and 56% shares, respectively, and most recently, deep learning-based approaches are under development.

Conclusion: The investigations show that using deep learning models in road network detection provides better results than traditional methods and gradually replaces these methods. Deep learning models with the ability to extract complex features and reduce the need for human intervention have improved the accuracy and efficiency of the detection process. On the other hand, super-resolution techniques based on deep learning can solve problems arising from the lack of high-resolution images by increasing the spatial resolution of images. By preserving spectral features and reducing noise, these techniques can provide higher-quality images for road detection. One of the main challenges in road detection from satellite images is the presence of vegetation cover and shadows, which can lead to incomplete and inconsistent detection of roads. To improve this problem, techniques such as tensor voting have been proposed, which can complete and correct roads that have been incompletely detected. Overall, combining super-resolution and deep learning methods for identifying road networks provides a cost-effective and efficient approach to updating road maps. These approaches, by reducing the costs and time required for detection, can be widely used by researchers and professionals in various fields. Furthermore, with further development and improvement of these techniques, new solutions can be developed for road construction, maintenance, and planning.