Leveraging sap flow data in a catchment-scale hybrid model to improve soil moisture and transpiration estimates

Sap flow encodes information about how plants regulate the opening and closing of stomata in response to varying soil water supply and atmospheric water demand. This study leverages this valuable information with model- data integration and deep learning to estimate canopy conductance in a hybrid catchment-scale model for more accurate hydrological simulations. Using data from three consecutive growing seasons, we first highlight that integrating canopy conductance inferred from sap flow data in a hydrological model leads to more realistic soil moisture estimates than using the conventional Jarvis-Stewart equation, particularly during drought conditions. The applicability of this first approach is, however, limited to the period where sap flow data are available. To overcome this limitation, we subsequently train a recurrent neural network (RNN) to predict catchment-averaged sap velocities based on standard hourly meteorological data. These simulated velocities are then used to estimate canopy conductance, allowing simulations for periods without sap flow data. We show that the hybrid model, which uses the canopy conductance from the machine learning (ML) approach, matches soil moisture and transpiration equally as well as model runs using observed sap flow data and has good potential for extrapolation beyond the study site. We conclude that such hybrid approaches open promising avenues for parametrizations of complex water-plant dynamics by improving our ability to incorporate novel or untypical data sets into hydrological models.