Resolving data gaps in global surface water monthly records through a self-supervised deep learning strategy

The distribution of land surface water bodies is constantly changing. Monitoring these changes is critical for both humanity and the ecological system. The Joint Research Centre Global Surface Water (GSW) dataset is crucial in monitoring global water resources. However, about a third of this dataset suffers from gaps and invalid observations, diminishing its practical effectiveness. We first highlight the challenge of filling data gaps in seasonal water areas, a task significantly more complex than for permanent water bodies. The distinction is vital, as most water bodies are permanent or exhibit high occurrence probability, simplifying gap-filling. Focusing on addressing this, we introduce a self-supervised learning approach tailored to address data gaps in seasonal water areas. The model, trained using simulated gaps from the GSW dataset, adeptly identifies water body seasonal fluctuation patterns. Tested against 639 cloud-free Sentinel-2 images and simulated labels, the model demonstrates high accuracy, achieving F1 score of 0.83 for water and 0.75 for land in seasonal water areas. To improve dataset reliability, we implemented a quality scoring system for each filled segment, distinguishing between high and low-quality filled data and substantially mitigating uncertainty. This study stresses the need for distinct validations for seasonal waters to circumvent biases inherent in methods that do not differentiate between water body types (permanent vs. seasonal). The gap-filled dataset enables a more precise and comprehensive analysis of global water resource trends, highlighting the transformative impact of deep learning in improving the utility and reliability of large-scale environmental datasets.