Using Machine Learning for Partitioning of Evapotranspiration into Evaporation and Transpiration in Flooded Ecosystems

Micrometeorological methods such as the eddy covariance method enable continuous measurement of ecosystem water vapor flux on time scales from individual half hours to years. However, they can generally only provide direct measurement of the combined net biosphere-atmosphere evapotranspiration (ET) flux (evaporation plus transpiration) above the plant canopy. The ability to partition micrometeorologically determined ET fluxes into transpiration and evaporation components would greatly improve our understanding of the water use of many ecosystems.

Most of the methods proposed for partitioning ET fluxes use the intrinsic relationship between carbon uptake and transpirational water loss, linked through stomatal exchange at the plant level, to estimate ecosystem transpiration. F looded ec osystems can have a significant contribution of evaporation from subcanopy water to total ET, which might violate some of the underlying assumptions made within partitioning methods based on the relationship between CO2 uptake and water loss.

We used Artificial Neural Networks (ANN) to separate ET from eddy covariance measurements into evaporation and transpiration at four different restored wetlands in the Sacramento/San Joaquin river delta of Norther n California .

We were able to predict realistic diurnal and seasonal changes in evaporation using ANNs, which combine nighttime evaporation measurements with environmental input variables such as air temperature, VPD and turbulence, to predict daytime evaporation . Transpiration fluxes were determined as the difference between total ET and predicted evaporation. The availability of ET data from a diverse set of wetland sites with varying maturities (time since restoration), canopy structure, and ratios of open water to vegetation cover provided an opportunity to test our partitioning method across a range of conditions, from evaporation dominated sites, such as during initial flooding of the wetlands, to transpiration dominated sites with a very dense canopy cover. By using data from different time periods and locations where transpiration fluxes were known or could be approximated we were able to validate our partitioning estimates in the context of our flooded wetland sites.