Comparative Analysis of Deep Learning Based Algorithms for Road Feature Extraction

Extraction of road features from remote sensing images is popular research and getting accurate road features has great practical applications such as urban planning, autonomous navigation, and disaster management. Accurate road feature extraction is challenging due to variations in road surfaces, occlusions, shadows, and complex environmental conditions. Recent advances in deep learning methods have shown great performance in road feature extraction compared with machine learning. In this paper, we evaluate and compare the performance of state-of-art road feature extraction including FCN, Unet, Segner, DeepLabv3+, and DCS-TransUpernet. This analysis is performed on the publicly available Deepglobe road extraction dataset, which consists of high-resolution satellite images. We explored the strengths and limitations of these models to understand the advancements made in road feature extraction and found that transformer-based models can capture more road information. This result sets a benchmark for future research in this domain.