Remotely sensed vegetation indices improve machine learning performance in predicting Fagus sylvatica L. forest transpiration in Mediterranean climate

Transpiration accounts for more than a half of the evaporative water fluxes from land. However, several criticalities still exist in its quantification over the whole land surface. Sap-flux measurement is a widespread technique that allows to retrieve data, at a high temporal resolution, for individual plants or trees. Unfortunately, due to practical constraints the spatial extent of sap-flux data is limited, and the campaigns of measurement are most often limited to one or few vegetative seasons. Hydro-meteorological data can be obtained in a much simpler way than sap-flux, while the vegetation condition is monitored frequently and at a high spatial resolution and with a wide coverage by several remote sensing satellite missions; both this kind of data are related to the transpiration process: on one hand the meteorological forcings drive the evaporation process, while the vegetation exerts control on the stomatal resistance, in response to both the environmental condition as well as its own physiological conditions. Machine learning (ML) is a suitable methodology for extracting and reproducing complex patterns from data; and thus might be able to predict sap-flux based on its physical drivers or proxies.

The objective of this research was to test three different ML algorithms (namely Regression Tree, Random Forest and XGBoost) on timeseries of transpiration based on sap-flux measurements taken in a Fagus sylvatica L. forest located in Southern Italy, during the 2021 and 2022 vegetative seasons, and to evaluate the performance of different vegetation indices (namely NDVI, EVI2 from Sentinel-2 and Cross-Ratio from Sentinel-1) in improving the prediction accuracy. As meteorological predictors Radiation, Air Temperature, Vapour Pressure Deficit, and Soil Moisture were selected. The accuracies obtained by training the algorithms on the meteorological dataset, were compared to those gained with the addition of the different vegetation indices.

The results showed that the vegetation indices always improved the prediction accuracy. EVI2 was the most effective vegetation index, and this is the first study to show that the Sentinel-1 Cross-Ratio is a valuable predictor of vegetation transpiration. With respect to algorithm performance Random Forest and XGBoost outperformed the Regression Tree and showed comparable accuracies between them.