Trends and variability of lake surface water storage in the source region of the Yellow River based on deep learning

The source region of the Yellow River is one of the main regions of the Asian water tower, lakes storing standing or slowly flowing water that provides essential ecosystem services of fresh water and food supply, waterbird habitat, cycling of pollutants and nutrients, and recreational services. Lakes are also key components of biogeochemical processes and regulate climate through cycling of carbon. Thus the estimation of trends and variability of the lake surface water storage is very critical, the direct human activities (damming and water consuption) and the natural factors (precipitation, runoff, temperature and potential evaporation) is gradually changing this environmentally sensitive region, especially the glacier retreat and permafrost thawing partially drive alpine lake expansion. The objective is mainly estimating the trends and variability of lake surface water storage using the deep learning module and long-term multi-source remote sensing data from the source region of the Yellow River. Optical remote sensing time-series images (Landsat 5-9, MODIS, and Sentinel-2) are employed to generate high-resolution, complete and closed lake surface shorelines and areas based on the Deep Convolutional Generative Adversarial Networks (DCGAN) deep learning method. Additionally, radar altimeters (GFO, T/P, Jason-1/2/3, Sentinel-6, ERS-2, Envisat, Cryosat-2, Saral/AltiKa, Sentinel 3/SRAL, ICESat, and ICESat-2) are utilized to recover lake water levels through the application of the Spatial-Temporal Graph Neural Networks (ST-GAN) deep learning method, providing insights into the long-term changes in lake water surface storage from 1992 to 2022. The study aims to assess the contributions of human activities and natural factors, and provids the valuable guidelines for water resource management.