Automated Road Extraction in Urban and Rural Regions from Remote Sensing Imagery with U-Net and ResNet-Based Architectures

Accurate road extraction from Remote Sensing Imagery (RSI) is crucial for various applications, including urban planning, navigation, and disaster response. This study investigates automated road extraction in rural, semi-urban and urban regions using deep learning architectures, specifically UNet and Deep Residual U-Net (ResNet) models. We leverage High-Resolution Satellite Imagery (HRSI) to train and validate our models, focusing on the distinct characteristics and challenges presented by urban and rural environments. Our methodology involves comprehensive data preprocessing and augmentation techniques to enhance model robustness. We evaluate the performance of both architectures using metrics and achieved an accuracy of 90.29%, recall of 80.95%, precision of 74.89%, and F1- score of 77.80% for the Deep Residual U-Net. Experimental results demonstrates that while both models perform well, the Deep Residual U-Net exhibits superior accuracy and generalization, particularly in complex urban landscapes.