GSWF: Global Surface Water gap-Filled dataset

The JRC Global Surface Water (GSW) dataset is a valuable resource for tracking changes in global water resources over 38 years (1984-2021), with a high-resolution of 30 meters. This dataset plays an essential role in understanding climate change, water management, and environmental monitoring. However, around one-third of the dataset is affected by gap or invalid observations, mainly due to issues with the Landsat archive. This data gap can significantly limit the dataset's usefulness in various real-world applications. To overcome this challenge, we developed a self-supervised learning method to fill in the data gap in seasonal water areas. Our deep learning model is trained to recognize the spatio-temporal patterns of water bodies and effectively fill in the data gap. We created a large training dataset using the JRC GSW dataset with simulated gap areas to teach the model how to fill in the gaps.Each GeoTIFF file in the dataset has two bands:1. Water/Non-Water (Binary): Indicates whether the pixel is water (1) or non-water (0).2. Good Quality/Bad Quality (Binary): Indicates whether the pixel is of good quality (1) or bad quality (0).The dataset uses a single-bit storage format to save memory.Quality Score CalculationThe quality score in our dataset is a binary score assigned to each 256x256 gap-filled block. It is calculated based on the proportion of valid observations within the block. A quality score of 1 indicates a high-quality block, while a score of 0 represents a low-quality block. The quality score calculation process is as follows:1. For each 256x256 block, count the number of valid observations (pixels with a "good quality" value of 1 in the Good Quality/Bad Quality band). Calculate the percentage of valid observations in the block by dividing the number of valid observations by the total number of pixels in the block (256x256 = 65,536).2. Assign a binary quality score to the block:If the percentage of valid observations is equal to or greater than 2%, assign a quality score of 1 (high quality).If the percentage of valid observations is less than 2%, assign a quality score of 0 (low quality).V2: Some data descriptions have been improved.