Comparing Machine Learning Models for Predicting Stomatal Conductance in 15 Urban Tree Species in Mexico City

Stomatal conductance (gS) is a key driver of urban tree transpiration and heat mitigation potential, but few studies compare machine learning models for predicting gS across multiple species in cities. This study applies five machine learning models (XGBoost, Random Forest, Support Vector Machine [SVM], Neural Network, and Random Forest Adjusted) and two classical models (Multiple Linear Regression and Generalized Additive Model [GAM]) to predict gS for 15 dominant tree species in the urban forest of Mexico City using environmental variables (air temperature, vapor pressure deficit, photosynthetically active radiation, and leaf water potential). We trained the models on a dataset of 300 observations per species, with 70% for training, 20% for validation, and 10% for testing, and evaluated performance using RMSE, MAE, and R2. Overall, XGBoost, GAM and SVM consistently showed the highest predictive performance, with R2 values up to 0.997, while the Neural Network and Multiple Linear Regression performed poorly (R2 ≈ 0.10–0.65). Model performance varied substantially among species, with XGBoost performing best for seven species, GAM for four, and SVM for four. Our results demonstrate that tree species gS can be accurately predicted using machine learning models in urban forests; however, model choice should account for species differences in performance. We therefore recommend that practitioners consider ensemble approaches of multiple models, excluding only the Neural Network, when selecting predictors for individual species.