Automatic Road Extraction using Scale Invariant Feature Transform-based Random Forest in Satellite Images

Road Extraction from High-Resolution Satellite Images (HRSI) plays a significant role in the analysis of the remote sensor image field. The extraction of road automatically labels a road’s pixels in a satellite image with semantic features depending on topographical meaningful features. However, manual road extraction inaccurately extracts road networks from satellite images which has significant difficulties like complex urban environments, varying road types, and diverse image conditions. Therefore, the Scale Invariant Feature Transform-based Random Forest(SIFT-RF) is proposed for automatic road extraction in satellite images. First, the image is obtained from DeepGlobe dataset and Z-score normalization is employed to maximize model performance. Then, SIFT is established to extract the road network in satellite images. Finally, the RF is utilized to extract and classify roads as road or non-road using Machine Learning (ML). The SIFT-RF achieves a greater f1-score of 98.24% compared to existing methods like Dual-Attention Road extraction Network (DA-RoadNet), High-Resolution U-Net architecture (HRU-Net), and Global Context-aware and Batch-independent Network (GCBNet) respectively.