Reconstruction of missing data in the global surface water dataset with a self-supervised deep learning model

This repository contains a subset of data for the paper "Reconstruction of Missing Data in the Global Surface Water Dataset with a Self-Supervised Deep Learning Model" (Under Review, Remote Sensing of Environment) for review by reviewers. The data includes reconstruction results for 50 reservoirs from the GRAND dataset, with each reservoir containing two parts: 1) reconstructions, and 2) paired images of JRC GSW, reconstructions, and differences (reconstructions minus JRC GSW) for each month for the convenience of reviewers. It is worth noting that this study does not reconstruct valid data (JRC GSW pixel values of 1 (land) or 2(water)) on the JRC GSW, and therefore, misclassifications on the JRC GSW are retained, sometimes resulting in visually incorrect images. The occurrence map (100_Occurrence.png) for each reservoir is provided with the maximum value of 100, meaning a pixel is always water across 440 months in the JRC GSW. The name of each folder (e.g. 00012) corresponds to the GRAND_ID in the GRAND dataset. Description of the color meaning in the comparison images (.png): In the JRC GSW and Reconstruction images, "red" represents missing values, "blue" represents water bodies, and "white" represents land. The "blue" in the Diff map represents newly identified water bodies, the "white" represents newly identified lands, and the "red" represents no change after reconstruction. Data Quality: The format of the file name is: YearMonth_Quality. Since the quality of the reconstruction is largely determined by the density of valid data available for reference in the current month, the quality of the reconstructed data is classified into three categories: "A": void data was less than 33%, "B": void data was between 33% and 66%, and "C": void data was greater than 66%.