Determining the Relative Influence of Water Potential, Biomass, and Temperature on Vegetation Optical Depth Using Physics-informed Machine Learning

Recent studies show that incorporating leaf water potential and plant hydraulics into land surface models can significantly improve evapotranspiration (ET) prediction. However, direct measurements of leaf water potential are destructive and very sparse. This largely limits their use to constrain large-scale plant hydraulics modeling. Meanwhile, vegetation optical depth (VOD), derived from microwave remote sensing, is often seen as a proxy for vegetation water content. Despite its wide applications for understanding water stress impacts on ecosystems, the relationships between VOD, leaf water potential and biomass still remain unclear. This gap has hindered our ability to use VOD to constrain large-scale land surface models. In this study, we develop a physics-informed machine learning model to predict VOD from water potential, leaf area index (LAI), temperature, and ecosystem attributes. The model is constrained by soil constitutive relations that convert soil moisture into water potential. Global remote sensing datasets of VOD (VODCA V2.0 and SMAP MT-DCA) and soil moisture (ESA CCI V8.1) are used to train the neural networks. We further apply the Explainable AI (XAI) technique, SHAP, to interpret how different input features (e.g. LAI, temperature, and water potential) contribute to VOD variability, and reveal how ecosystem attributes impact the water potential-VOD relations. This approach enables us to systematically examine the spatial variations of VOD-biomass-water potential relationships and the critical roles of ecosystem attributes in modulating these patterns. The results can enhance the applicability of VOD assimilation into land surface models, and thereby further improve the representation of plant hydraulics and ecosystem functions in large-scale models.