Deep Learning (CNN) for Detecting Road Infrastructure in Old Mosul City Using High-Resolution Aerial Imagery

Abstract

Road networks and transportation infrastructure play a crucial role in many applications such as urban planning and environmental assessment. Remote sensing provides multispectral data that can be used to identify, extract, and map roads. However, accurate mapping of road networks from aerial imagery poses challenges due to the complexity of real-world road patterns. The aim of this study was to develop deep learning techniques for automated road extraction from aerial photographs. The study evaluated several convolutional neural network (CNN) architectures were including a hybrid spectral-spatial network (HybridSN). Models were assessed on a dataset of urban aerial images with lidar-derived ground truth labels. The joint modeling of multi-modal cues enables highly precise localization and delineation of road segments. The results of the HybridSN integrating both spectral and spatial processing achieved the top performance with 96.9% overall accuracy and 80.6% intersection-over-union after post-processing. In comparison, CNNs leveraging spatial context alone perform worse with the best overall accuracy of 95.4% after post-processing. The findings demonstrate the importance of fusing spectral and spatial data within deep learning frameworks for road extraction.