The Construction of 20‐year Daily Surface Albedo Along PANDA Transect, Antarctica

Abstract Surface albedo, which represents the Earth's capacity to reflect solar radiation, is critical for energy balance, climate modeling, and weather forecasting. However, ground observation stations are sparsely and unevenly distributed in Antarctica, and reanalysis and satellite data often exhibit errors, making it challenging to accurately monitor surface albedo. This study integrates nine automatic weather station data along PANDA transect, ERA5‐Land reanalysis, and MCD43C3 product, utilizing a dual‐layer stacking ensemble approach that incorporates various machine learning algorithms, including random forest (RF), light gradient boosting machine (LightGBM), multilayer perceptron (MLP), extreme gradient boosting (XGBoost), and gradient boosting decision tree (GBDT). The results indicate that stacking model achieves the best accuracy along the 1,200 km PANDA transect from Zhongshan Station to Dome A, demonstrating high correlation (Corr = 0.939) and low error metrics (MSE = 0.004, MAE = 0.047, RMSE = 0.064). Additionally, optimal models were developed individually for each station, which provided accurate daily albedo predictions from 2004 through 2023 for nine stations in East Antarctica. Compared to albedo products such as ERA5‐Land (Bias = 0.1217) and MCD43C3 (Bias = 0.0591), the data set constructed demonstrates higher accuracy with a Bias of 0.0041, underscoring the superior generalization of the machine learning models. Based on these high resolution data sets at the nine sites, it is found that the surface albedo from coast to Dome A exhibits a declining trend during 2004–2023, indicating an ongoing darkening process. This phenomenon reflects the important response and feedback of surface snow to global climate change, and the high‐quality long‐term data set may provide deep insights in the future application.