Road extraction from satellite images using Deep Learning on HPC

Abstract. Road extraction from satellite imagery is required for various applications, such as infrastructure planning, city development, transportation, and disaster response planning. Road extraction is treated as semantic segmentation problem where U-Net is proven to be highly effective. However, the standard U-Net model consists approximately 31 million parameters, which leads to high computational costs and resources. To address this challenge, we have modified U-Net architecture by integrating Depth-wise Separable Convolutions (DSC) that reduced the number of trainable parameters down to 3.5 million and improved the model ability to capture road features. For this work, we utilized a combination of open-source satellite imagery datasets and a custom dataset. This dataset contains 53877 high-resolution satellite image and road mask pairs. Comparative result shows that the U-Net with DSC achieves a mean accuracy of 0.956 and a mean Intersection over Union (mIoU) of 0.626, outperforming the standard U-Net, which recorded a mean accuracy of 0.949 and mIoU of 0.581. These models are trained on C-DAC PARAM Siddhi-AI high-performance computing (HPC) infrastructure. The prediction results improve the road continuity and clarity, making the model suitable for national initiatives such as PMGSY National GIS. This model integration can greatly enhance the efficiency of planning, monitoring, and execution of road infrastructure projects.