PLA4MS: Curated Georeferenced Dataset for Cloud Removal in Remote Sensing

The presence of meticulously curated extensive training datasets plays a crucial role in advancing the performance of deep learning techniques that generalize well for extracting geoinformation from multi-sensor remote sensing imagery. Despite numerous datasets being published by the research community, a substantial portion of them is hampered by significant constraints, such as low spatial resolution, lack of ground details over time, and insufficient sample quantity. The paper utilizes the openly accessible data obtained from the Planetscope satellites managed by Planet Labs. We have curated a dataset called PLA4MS comprising 64,557 pairs of images depicting cloudy and cloud-free conditions. The study focuses on the Ropar region of Punjab, India, as the primary area of interest, ensuring precise georeferencing at a spatial resolution of approximately 3 meters across all meteorological seasons. This work presents a comprehensive cloud-removal dataset aimed at advancing remote sensing techniques, with cloud removal as the primary focus. This dataset is created to serve the research community, as it offers the potential to support broader remote sensing applications by enabling researchers to generate cloud-free images for time-series analysis, land cover land use classification, change detection, and other agricultural tasks.