Transpiration in forest ecosystems based on deep learning and sap flow observations

Transpiration (T), the component of evaporation (E) controlled by vegetation, dominates terrestrial Evaporation, but measurements are highly uncertain. In the light of the importance of evaporation for studying the terrestrial water cycle, hydro-climatic extremes such as droughts and heatwaves and land-atmospheric interactions, there is a strong demand on novel approaches to reliably estimate T. Currently available approaches to estimate T mostly rely on its relationship with photosynthesis, but parameterizing this relationship is difficult and estimates of T strongly disagree among each other in terms of magnitude. Moreover, in-situ measurements are scarce and and evaporation cannot be measured directly from space.

We developed a hybrid Priestley-Taylor (PT) model using Deep Learning to learn the relationship between T and state variables such as soil moisture, vapor pressure deficit and the fraction of photosynthetic active radiation for different plant functional types (PFTs). We use globally available variables from reanalysis and remote sensing data as forcing to train an artificial neural network on the PT-coefficient α obtained by inverting the PT model on sap flow based ecosystem T. In this way, we can predict Transpiration at local scales independently from hard-to-obtain fluxes like E or vegetation parameters such as stomatal conductance. We evaluate our algorithm against T estimates from flux partitioning methods based on water use efficiency at eddy covariance sites for different PFTs and regions. Also, we compare our estimates with other available products of transpiration like GLEAM, PML-V2 and ERA5-Land. Preliminary results of this research showed that the developed model can learn the relationship between T and few influencing variables, without incorporating variables such as net radiation or GPP. Our findings contribute to dissolving the scarcity of T estimates in forest ecosystems based on actual observations. Future work is needed to apply our method to the larger scale for studying spatial patterns of T, e.g. across the European continent.