Evapotranspiration Partitioning Using Machine Learning

Evapotranspiration from terrestrial ecosystems plays a crucial role in the global water cycle. Understanding the relative contribution of transpiration (T) and evaporation (E) is essential to improve our understanding of the water cycle and our ability to predict its response to climate change. While measurements of the combined flux of E plus T (evapotranspiration, ET) are numerous, information about the relative contribution of T versus E is not readily available at the ecosystem scale. This is particularly true for ecosystems with a relatively large contribution of E, such as wetlands.

We have developed a methodology for partitioning of micrometeorological ET measurements using Machine Learning (ML) tools to predict E from nighttime ET measurements. A robust comparison of machine learning models included: parametric models, as they are fast and improve performance by adjusting parameter weights; non-parametric models, which both build a predictive function and determine its parameter values from within the data; and ensemble methods which combine the predictions of many simple models to produce the most accurate predictive model. This method was tested and validated in wetlands (Eichelmann et al., 2022; Stapelton et al.; 2022) and its applicability to select forest ecosystems is investigated and compared to alternative partitioning methods. Using this framework to identify the best performing machine learning algorithm across a number of sites has revealed that there is no single optimal machine learning algorithm and that site specific algorithm testing and selection is recommended.

However, it was also recommended that deeper insights can and should be facilitated when using a machine learning approach. In order to both deliver these insights and identify any previously hidden correlations from the significant volume of ET data available, we employed a detailed feature importance study. This delivered a ranking of features for each site where the feature set is optimised by a recursive feature elimination algorithm. Identifying new (important) features may reveal previously unknown connections between components of the system for further study with the potential to improve understanding of the underlying biophysical processes.

Eichelmann, E., Mantoani, M. C., Chamberlain, S. D., Hemes, K. S., Oikawa, P. Y., Szutu, D., Valach, A., Verfaillie, J., & Baldocchi, D. D. (2022). A novel approach to partitioning evapotranspiration into evaporation and transpiration in flooded ecosystems. Global Change Biology, 28, 990– 1007. https://doi.org/10.1111/gcb.15974

Stapleton, A., Eichelmann, E., & Roantree, M. (2022). A framework for constructing machine learning models with feature set optimisation for evapotraspiration partitioning. Applied Computing and Geoscience, 100105. https://doi.org/10.1016/j.acags.2022.100105