Mapping snow-covered forest albedo via hybrid radiative transfer and machine learning across Northern Hemisphere

Abstract

Snow-covered forests are widely distributed in the Northern Hemisphere, and their albedo significantly influences forest warming effects, radiative balance, and climate change. However, estimating snow-covered forest albedo is challenging due to the complex interactions between the snow and canopy. Current algorithms often rely on snow-free forest models or linear weighting of snow and forest components. These simplified forward models result in significant errors in the bidirectional reflectance simulation of snow-covered forests. Meanwhile, the albedo retrieval process is computationally intensive, especially when lookup tables or optimization algorithms are employed. Thus, we propose a novel albedo retrieval framework that integrates the strengths of snow-covered forest radiative transfer model with the efficiency of machine learning methods. This framework achieves three key advancements: (1) the snow-covered forest bidirectional reflectance (SFBR2) model is extended to sloped terrain to reduce the reflectance simulation errors; (2) the representativeness and accuracy of training datasets are improved by combining satellite observations with SFBR2-retrieved albedo; and (3) Random Forest model is utilized on the Google Earth Engine (GEE) platform to enable rapid retrieval of snow-covered forest albedo. As a result, a snow-covered forest albedo product for the Northern Hemisphere from 2001 to 2022 is successfully generated. Validation against albedo observations from flux stations demonstrates that our retrieval framework achieves higher accuracy (R² ≥ 0.775 and RMSE ≤ 0.037) than the Moderate Resolution Imaging Spectroradiometer (MODIS) and Global LAnd Surface Satellites (GLASS) products. This highlights its potential to further enhance our understanding of radiative balance and climate change in snow-covered forests.