Predicting evapotranspiration using machine and deep learning methods

Brenner, Claire; Frame, Jonathan; Nearing, Grey; Schulz, Karsten

Predicting evapotranspiration using machine and deep learning methods | Schätzung der Verdunstung mithilfe von Machine- und Deep Learning-Methoden

2021

Brenner, Claire | Frame, Jonathan | Nearing, Grey | Schulz, Karsten

Evapotranspiration is a key driver of the Earth’s climate system, modulating processes in the global water, energy as well as carbon cycle. Information on evapotranspiration rates and its variability in space and time are therefore essential for climate modelling, examining the effect of climate change and sustainable agricultural management.In this study we apply two machine- and deep learning methods for predicting evapotranspiration at half-hourly and daily temporal resolutions using data from the FLUXNET network. We train a Long Short-Term Memory, a recurrent neural network with explicit memory that is particularly suited for time series predictions (similar to physically-based water balance models). As a second model we use XGBoost, a regression tree method that does not contain system memory (similar to physically-based energy balance models). With this setup we want to examine the importance of memory effects for predicting evapotranspiration.Our analyses show that both modelling approaches agree well with the observations. Averaging over all 153 studied FLUXNET sites, the LSTM performs significantly better than XGBoost. However, model performance varies with vegetation class; the LSTM yields better results for warmer and drier sites with low vegetation, whereas XGBoost performs better for e.g. wetland sites. Thus, the importance of system memory seems to vary with ecosystem and climate type.The presented results underline the potential of artificial intelligence for predicting evapotranspiration.

Afficher plus [+]