Optimising soil-hydrological predictions using effective CART models
There are various problems with process-based models at the landscape scale, including substantial computational requirements, a multitude of uncertain input parameters and the limited parameter identificability. Classification And Regression Trees (CART) is a recent data-based approach that is likely to yield advantages both over process-based models and simple empirical models. This non-parametric regression technique can be used to simplify process-based models by extracting key variables, which govern the process of interest at a specified scale. In other words, the model complexity can be fitted to the information content in the data. CART is applied to model spatially distributed percolation in soils using weather data and the groundwater depths specific to the site. The training data was obtained by numerical experiments with Hydrus1D. Percolation is effectively predicted using CART but the model performance is highly dependant on the available data and the boundary conditions. However, the effective CART models possess an optimal complexity that corresponds to the information content in the data and hence, are particularly suited for environmental management purposes.