Advancing Global Land Surface Albedo Parameterization with Physics-Informed Machine Learning Methods

Land Surface Albedo (LSA) directly affects the Earth's surface energy budget, exerting substantial influence on climate and environmental dynamics, including global warming, snow and ice melt, vegetation and soil degradation, and urban heat islands. Considering the intricate relationship between LSA, incident solar radiation, and sub-grid surface characteristics, along with the constraints of limited and localized observational data, there arises a necessity to parameterize these interactions using satellite observations and model calculated variables within climate models. The primary objective is to improve the parameterization and markedly reduce model-induced biases in surface albedo predictions, leading to enhanced performance of Global Climate Models (GCMs) and improved climate predictions. This study presents a global LSA parameterization scheme utilizing a Physics informed-machine learning (ML) approach. Our methodology integrates satellite data, land surface model (LSMs) outputs, and reanalysis data spanning a decade, covering diverse surface types such as soil, snow, and vegetation. The physical component captures variations in albedo, leveraging known dynamical relations with variables including solar zenith angle, surface soil moisture, fractional vegetation cover, leaf and stem area index, greenness, topography, soil texture, snow fraction, snow impurities, grain size, and age. Simultaneously, the ML-based model optimizes the scheme's parameters by closely adhering to the underlying dynamic relationships and physical constraints. Concurrently, a statistical component of the ML model corrects static effects specific to local surface characteristics. This study stands out for its comprehensive global coverage at high spatial and temporal resolutions, enabled by ML models.