Road extraction in diverse urban environments using UAV data and nDSM perturbations: A case of Bhopal, India

Abstract

Automated road extraction has a wide range of applications in urban planning, transportation management, and emergency response. However, existing methods struggle to extract roads in dense regions of developing countries, where the road networks are diverse and unplanned. This is due to the common spectral signatures between roads and neighboring objects, as well as the limited ability of current methods to combine multispectral and RGB images with normalized digital surface models (nDSMs). To address these challenges, we propose a novel approach that integrates UAV imagery from the Gehukheda region in Bhopal, India with high-resolution elevation data obtained from generated nDSMs and leveraging multispectral (RGB and NIR) and true-color RGB images to differentiate materials and elevation differences. We also introduce feature-aware strategic perturbations in the nDSM to improve segmentation efficiency. We trained three deep learning models, VGG19-UNet, DeepLabV3+, and SegFormer-B5 on our manually labeled training data. All three models performed well with the incorporation of nDSM and NIR. The perturbed DSM provided significantly better results, increasing the overall IoU of roads from 90.95% to 92.16% for VGG19-UNet, 90.59%–91.29% for DeepLabV3+, and from 91.75 to 93.68% for SegFormer-B5. These results demonstrate the effectiveness of our proposed approach in accurately segmenting roads, particularly within dense informal settlements. The proposed approach can help to overcome the limitations of satellite imagery and existing road extraction methods, thereby enhancing the accuracy and efficiency of road network identification and analysis in densely populated urban environments of developing countries.