Road Network Extraction from Satellite Images Using Deep Learning

Abstract

AbstractRoad network extraction plays a vital role in traffic analysis and safety monitoring, as well as in analysing, designing, and maintaining road structures. The task of road network extraction is tedious due to occlusions, shadows, non-road objects, and diversity of road network structures such as gravel, asphalt, sand. This study investigates road network extraction using deep learning techniques on high-resolution satellite images. More specifically, the combination of DenseNet and UNet deep neural networks is investigated for proficiency in road network extraction using high-resolution satellite images. Combining these complex neural networks has the potential to allow a deeper extraction of characteristics of a road network attribute, thereby increasing the accuracy in detecting and extracting all types of road networks. In addition, the use of a large dataset with very high-resolution images is to train the model, further increasing the accuracy of the model. The final combined neural network UNet-DenseNet-UNet, coupled with the high-resolution images, helps generate results that were better when compared to existing comparable models in literature which use deep learning techniques, with the intrinsic difference in the dataset used being the resolution of the images used to train the model.KeywordsRoad network extractionDeep learningMachine learningUNetDenseNet