Robust Machine Learning-Based Albedo Parameterization: Generalization Across Land Cover Schemes

Land Surface Albedo (LSA) significantly influences the Earth's surface energy budget, playing a crucial role in climate and environmental dynamics. Given the complex interplay between LSA, incident solar radiation, and subgrid surface characteristics, alongside the challenges posed by limited and localized observational data, there is a critical need to parameterize these interactions using satellite observations and model-derived variables in climate models. To address this, we have developed a global LSA parameterization scheme using a physics-informed machine learning (ML) approach. Global and regional climate models utilize various land cover classification (LCC) schemes, such as IGBP, USGS, UMD, and PFT. These classification systems differ significantly, making direct conversion between them complex and non-trivial. Additionally, some of these datasets are global or regional, while others are time-varying or remain static. This variability poses challenges for LSA parameterization efforts that rely on consistent classification schemes. In this study, we present and compare results for three different LCC schemes: the MODIS IGBP 16-class, USGS 24-category, and long-term MODIS IGBP 16-class climatology. Our findings demonstrate the adaptability of the proposed ML methods, which can be effectively modified and trained to accommodate different LCC schemes. This adaptability makes ML a powerful tool for generalizing the LSA parameterizations for integration into Regional Climate Models (RCMs), Global Climate Models (GCMs), and Land Surface Models (LSMs) with varying LCC specifications. The parameterization scheme we developed shows consistent performance across different LCC schemes, underscoring the robustness of the features and the ML architecture's ability to capture albedo dynamics. This approach offers a more accurate representation of the variability and complexity of LSA, providing a significant improvement over traditional methods. These results highlight the potential of ML-based parameterizations to enhance the accuracy and applicability of climate models, irrespective of the LCC scheme employed.