A long-term surface albedo dataset for multiple sites from Zhongshan station to Dome A, East Antarctica (2004-2023)

Surface albedo represents the Earth's ability to reflect solar radiation and plays a crucial role in areas such as land surface process models, surface-atmosphere radiation energy balance, climate modeling, weather forecasting, and disaster monitoring. This dataset encompasses nine automatic weather stations, including Zhongshan, Panda 100, Panda 300, Grove 01, Panda 400, Taishan, Panda 700, Panda 1100, and Kunlun Station, providing daily predicted surface albedo values from 2004 to 2023 in East Antarctica. The model integrates observational data, ERA5-Land reanalysis data, and the MODIS albedo product, employing advanced machine learning models, including RF, LightGBM, MLP, XGBoost, and GBDT, with a stacking model as the meta-model to achieve more accurate albedo simulations. The Panda 300 and Panda 700 sites showed optimal performance with the RF model, while other sites demonstrated higher accuracy with the stacking model. Except for Zhongshan Station, a correlation coefficient above 0.91 and an MSE less than 0.08 were recorded at all other sites. Compared to other albedo products, such as ERA5-Land, ERA5, and MCD43C3, the machine learning models produced more accurate simulations with lower errors, fully demonstrating their generalization and transfer learning capabilities. Rigorous quality control measures, along with approximately 10,000 training iterations and multiple debugging sessions, ensured the generalizability and robustness of the model. Essential for exploring the spatiotemporal variations of surface albedo, this dataset offers invaluable insights and robust data support for understanding the Antarctic surface-atmosphere feedback mechanisms and their implications for global climate change.