Sentinel-2based dataset covering Indian regions for road detection: A dataset for infrastructure monitoring

Road extraction from satellite imagery plays a vital role in infrastructure monitoring applications such as transportation planning, autonomous navigation and disaster response. Unlike very high-resolution (VHR) imagery, Sentinel-2 data is publicly available, thereby being a strong contender for remote sensing problems such as road extraction. We present the Sentinel-2 Roads Dataset, a publicly available, open-access dataset designed for road segmentation tasks using Sentinel-2 satellite imagery. The dataset consists of Sentinel-2 images at 10 m spatial resolution acquired from seven diverse regions across India, covering urban, rural, and mountainous terrains. The corresponding binary road masks were generated using OpenStreetMap (OSM) data and were rasterized using QGIS software and Rasterio library. The dataset also offer a set of enhanced version of the Sentinel-2 images, which incorporates techniques such as gamma correction and CLAHE, for improved road visibility. For ease of reuse in different deep learning frameworks, the dataset is made available in TIF as well as PNG format. The dataset contains a total of 5634 high-quality samples, each of 256 × 256 pixels, obtained after manual refinement. As a baseline, several deep learning segmentation models were trained on both raw and enhanced images, which demonstrated that preprocessing techniques significantly improved segmentation performance . The proposed dataset serves as a valuable resource for computer vision and remote sensing researchers, offering a benchmark for road extraction tasks. It is publicly available for further research and development in road segmentation models using satellite imagery.