Publication

Abstract : Generalising UNet-based models for road extraction from high-resolution remote sensing datasets remains challenging in practical applications. This work evaluates seven UNet variants using the Massachusetts Roads, DeepGlobe, and CHN6-CUG datasets. To test generalization ability we evaluated their transferability to our Unmanned Aerial Vehicle (UAV) imagery. We apply conventional evaluation criteria such as IoU, precision, recall, F1-score, and computing cost. The results demonstrate a consistent trend: higher architectural complexity produces relatively minimal improvements. Also, each model exhibits a notable drop in performance when evaluated with images from other road dataset, highlighting that cross-dataset adaptation continues to be an important challenge. The original UNet demonstrates comparable performance to more advanced models while having significantly more computational efficiency, thereby proving its position as a solid baseline. The results highlight the need for segmentation models that can adapt to the significant variability found in real-world data.