Data-driven prediction of urban vegetation cooling effects using machine learning and field observations

The rapid expansion of urban populations, coupled with growing epidemiological evidence that associates extreme temperature events with adverse health outcomes and elevated mortality rates underscores the critical role of Urban Green Areas (UGAs) in delivering ecosystem services that enhance human well-being. Among these services, the air temperature cooling potential (ΔT°C) driven by ecosystem evapotranspiration (ET) represents a key mechanism for mitigating heat-related health risks.

This study investigates the capacity of various Machine Learning (ML) algorithms to predict the ΔT°C of UGAs, thereby supporting thermal regulation through ET and highlighting their importance in sustainable urban planning and climate adaptation strategies. We used multiple years of experimental Eddy Covariance (EC) observations of the ET the to train and validate a series of ML algorithms with the objective to simulate the cooling effect of the urban vegetation. A preliminary analysis of predictor variables was conducted to identify and rank their importance using the mean absolute Shapley (Sh) values. Results indicate that incoming shortwave solar radiation (Rg) was the most influential predictor (Sh = 0.45), followed by vapor pressure deficit (VPD, Sh = 0.20), relative humidity (RH, Sh = 0.075), air temperature (AirT, Sh = 0.065), friction velocity (u*, Sh = 0.02), and wind speed (WS, Sh = 0.01). The application of ML algorithms revealed that Bootstrap Aggregation (Bagging) and Least-Squares Boosting (LSBoost) performed best, achieving R² values of 0.89 and 0.83, respectively, during the training phase compared to observed data. Other algorithms, including Neural Networks (NN), Gaussian Process Regression (GPR), and Support Vector Machines (SVM), showed also similar, but slightly lower r² , with values ranging from 0.80 (NN) to 0.79 (SVM). Ten-fold cross-validation confirmed robust generalization, as model performance remained consistent regardless of the data subset used to compute R² between modeled and observed values. Further evaluation using Taylor diagrams showed that the average normalized standard deviation (σn) and Pearson correlation coefficient of the models were 0.89 (±0.02) and 0.90 (±0.02), respectively, closely matching the observed data.

During the testing phase we observed, as expected, a clear reduction of the ML performance compared to the training phase: however, over the three years of the testing phase, RG bagging and LSBOOST have confirmed their superiority, compared to the other algorithms, with an average r² between observed and simulated data of 0.66 and 0.67 respectively. Discrepancies between predicted and observed ΔT°C during testing were most evident during midday hours, with an average overestimation of 0.31°C (±0.2).

Overall, the investigated UGAs demonstrated an average capacity to reduce ambient air temperature during summer by approximately 2°C to 4°C.