Abstract Accurate surface albedo parameterization is critical for modeling Earth's energy balance. Yet, many schemes rely on static look‐up tables or semi‐empirical formulations that fail to capture spatiotemporal variations and complex radiative interactions. This study develops a physics‐informed machine‐learning parameterization using 19 years (2003–2021) of MODIS Bidirectional Reflectance Distribution Function data to predict direct and diffuse albedo in the visible and near‐infrared bands across major land‐cover categories. The framework leverages 10 biogeophysical predictors, including solar geometry, vegetation state, soil moisture, soil texture (STX), topography, and background climate. It comprises a dynamic component that resolves physical spatiotemporal patterns and a static correction, the Surface Albedo Localization Factor, which accounts for subgrid heterogeneity, together explaining most of the observed variability. The parameterization shows strong agreement with MODIS albedo (overall R 2 = 0.88, mean absolute percentage error MAPE = 7.5%), with performance ranging from R 2 = 0.85 in direct visible to R 2 = 0.90 in diffuse near‐infrared, and generally higher accuracy in the near‐infrared parts. It performs well across diverse LCCs, including grasslands, shrublands, croplands, and challenging barren regions where empirical methods underperform. SALF improves accuracy across all albedo parts (average R 2 increase of 0.16 and MAPE reduction of 5.1%). Feature‐importance analysis indicates that solar zenith angle and leaf area index are the dominant drivers of dynamic variability, whereas STX and topography influence static variability. Counterfactual experiments confirm biophysically consistent albedo responses, enhancing interpretability and model trust. This framework offers a physically grounded alternative to empirical schemes and has strong potential for integration into Earth system models to improve the representation of surface energy exchange.
Ralhan et al. (2026) studied this question.