Towards improved simulation of water partitioning: which observations have most value in constraining a spatially-distributed ecohydrological model?

Improved quantification of water partitioning is needed to inform sustainable, integrated land and water management. Spatially-distributed, process-based ecohydrological models are promising tools for achieving this; however, such models typically have many parameters that require estimation from data. Utilising an extremely rich plot-scale dataset incorporating energy (net radiation, temperature, and latent and sensible heat), hydrological (soil moisture, sap flow and actual evapotranspiration) and vegetation (net primary productivity and structure) components, we investigated which types of observation best constrain the parameters of a complex ecohydrological model (EcH₂O-iso) for simulation of water partitioning. Our experimental site was situated within one of the largest coffee agroforestry systems in Costa Rica and experiences high energy inputs and intense rainfall events, thus adding further complexity to the robust simulation of hydrological processes. A series of calibration exercises were undertaken based on combinations of the different observation types. In each case, 100 behavioural parameter sets were chosen following 100,000 Monte Carlo simulations. The “flux mapping” approach was then used to quantify the percentage contribution made by different simulated fluxes to total model outflows (e.g., contributions of transpiration, soil evaporation and interception evaporation to total evapotranspiration), in order to assess how consistently plot hydrology was simulated by the retained parameter sets. Additionally, PCA analysis of performance metrics (including those for observations not used in the given calibration) was undertaken to reveal how contrasting observation types “pull” the model in different directions and, thus, affect its ability to capture the dynamics of each type simultaneously. From this work, we are able to provide guidance on how different ecohydrological datasets may be optimally combined in model calibration. This has implications not only for reducing uncertainty in modelling studies underpinning land and water management, but also for designing future field campaigns such that collection of the most valuable data can be prioritised.