Evaluation of UNet architecture variants for road network extraction from high resolution geospatial dataset

This paper assesses the generalization capacities of UNet variants for road extraction from satellite imagery. The models were trained on three benchmark datasets: MIT Roads Dataset, DeepGlobe Road Extraction Dataset, and CHN6-CUG Road Dataset. To evaluate their performance, the trained models were subjected to Unmanned aerial vehicle imagery from the Thiruvallam area, a region that is distinguished by its diverse land cover types and intricate geographic extent. The research uncovers the advantages and drawbacks of each UNet variation in terms of intersection over union, precision, recall, F1-score, and computational efficiency. The findings highlight the capabilities of modern UNet architectures for robust road extraction from various types of aerial images.