Transpiration is a key biogeochemical process, accounting for more than half of the evaporative water fluxes from land to the atmosphere; however, its quantification is still a hot topic. Sap-flux is a commonly used technique to measure the transpiration of individual plants or trees at a high temporal resolution but limited in time and space to the measurement campaigns. The quantification of hydro-meteorological parameters, (e.g. air temperature, incoming radiation, soil moisture etc.) that drive the transpiration process, is way simpler. The condition of vegetation, which influences transpiration by modulating the stomatal resistance, is extensively monitored by several remote sensing satellite missions. Three different Machine Learning (ML) algorithms (Regression Tree, Random Forest and XGBoost) are tested on the 2021 and 2022 timeseries of sap-flux based transpiration measured in a Fagus sylvatica forest located in Southern Italy, to evaluate the usefulness of different vegetation indices (namely NDVI, EVI2 from Sentinel-2 and Cross-polarization Ratio (CR) from Sentinel-1) in increasing the prediction accuracy. As meteorological predictors Radiation, Air Temperature, Vapour Pressure Deficit, and Soil Moisture were selected. ML was chosen due to its effectivity in extracting the complex and non-linear interplays between predictors and the response variable. The results showed that the inclusion of vegetation indices in the predictors always improved the prediction accuracy. EVI2 was the most effective vegetation index, and this is the first study to show that the Sentinel-1 CR 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. The added value of Cross-Ratio is that, being sensed in the Radar wavelength, it is not affected by the atmospheric conditions, and thus might be helpful in areas that experience significant cloud cover. Our findings show different suitable approaches for upscaling sap-flux timeseries, depending on the context of application and useful to reconstruct forest transpiration at local and regional scale.
Reconstruction of the dynamics of sap-flow timeseries of a beech forest using a machine learning approach
Kabala J. P.;Niccoli F.;Battipaglia G.
2025
Abstract
Transpiration is a key biogeochemical process, accounting for more than half of the evaporative water fluxes from land to the atmosphere; however, its quantification is still a hot topic. Sap-flux is a commonly used technique to measure the transpiration of individual plants or trees at a high temporal resolution but limited in time and space to the measurement campaigns. The quantification of hydro-meteorological parameters, (e.g. air temperature, incoming radiation, soil moisture etc.) that drive the transpiration process, is way simpler. The condition of vegetation, which influences transpiration by modulating the stomatal resistance, is extensively monitored by several remote sensing satellite missions. Three different Machine Learning (ML) algorithms (Regression Tree, Random Forest and XGBoost) are tested on the 2021 and 2022 timeseries of sap-flux based transpiration measured in a Fagus sylvatica forest located in Southern Italy, to evaluate the usefulness of different vegetation indices (namely NDVI, EVI2 from Sentinel-2 and Cross-polarization Ratio (CR) from Sentinel-1) in increasing the prediction accuracy. As meteorological predictors Radiation, Air Temperature, Vapour Pressure Deficit, and Soil Moisture were selected. ML was chosen due to its effectivity in extracting the complex and non-linear interplays between predictors and the response variable. The results showed that the inclusion of vegetation indices in the predictors always improved the prediction accuracy. EVI2 was the most effective vegetation index, and this is the first study to show that the Sentinel-1 CR 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. The added value of Cross-Ratio is that, being sensed in the Radar wavelength, it is not affected by the atmospheric conditions, and thus might be helpful in areas that experience significant cloud cover. Our findings show different suitable approaches for upscaling sap-flux timeseries, depending on the context of application and useful to reconstruct forest transpiration at local and regional scale.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
