A Dual Model Approach to Better Generalize Individual Tree Water Use

Transpiration is a critical component of the carbon-water cycle, driving water from the soil to the atmosphere through plants as sap flow and linking plants to larger climate fluctuations. While the measurement of sap flow using thermometric principles has been refined over decades, translating these in-situ measurements into generalized models remains a challenge. The availability of the SAPFLUXNET database now opens opportunities for global, data-driven modeling. Despite the sophistication of recent approaches, current models often demonstrate high accuracy within specific sites but suffer performance degradation during cross-site validation. This study aims to overcome the generalization gap by introducing a modeling framework that decouples the prediction of temporal dynamics from absolute magnitude using a dual-model approach. The first model predicts normalized temporal patterns based on the 90^th percentile and nighttime sap flow for each tree using XGBoost. The second model predicts absolute magnitude using tree and site-level characteristics using Random Forest. Results show a significant improvement in overall performance, with R² increasing from 0.47 to 0.54 compared to a single combined model. This gain is primarily driven by better performance in the temporal model. While the average Root Mean Squared Error (RMSE) showed only minor improvement, the performance gains were consistent across tree sizes, genera, and plant functional types, validating the dual-model approach. Future work could further improve this framework by incorporating memory-based temporal models and integrating trait and remote sensing datasets for better tree representation. Finally, this scalable approach can be adopted to estimate regional-scale transpiration using species and tree size distributions, helping to refine our understanding of tree water use.