Inverse modelling with a genetic algorithm to derive hydraulic properties of a multi-layered forest soil

Abstract

Modelling complex vadose-zone hydrological processes across a range of spatial scales requires models and hydraulic properties commensurate with the scale of investigation. This study investigates the ability of two conceptual models with contrasting complexity and parameter requirements to quantify accurately the one-dimensional water balance in a soil–vegetation–atmosphere (SVA) system. The two models tested were: (i) the mechanistic HYDRUS-1D model, which numerically solves the Richards equation for saturated–unsaturated water flow; and (ii) a compartment or budget model that includes simplified schemes for redistribution of water in the soil. We discuss model performance for parameter sets obtained by inverse modelling for an SVA system developed in a podzol soil with Scots pine vegetation in Belgium. Soil hydraulic properties were derived from field-based soil water content data collected at multiple depths in two lysimeters installed in the multi-layered forest soil and subject to atmospheric boundary conditions during nearly one full hydrological year. Parameter optimisation was based on a genetic algorithm including elitism as an operator for improving the search for optimal solutions with better performance scores. Four scenarios were developed to investigate (i) the impact of the type of conceptual flow model (mechanistic or compartment), and (ii) the effect of the degree of detail or granularity used to describe the soil profile, on the accuracy of inverse modelling (i.e. five or two material layers with different hydraulic properties or a homogeneous profile with effective properties). Results showed that for models with the same number of material layers as the number of pedogenic horizons in the soil profile, both conceptual models reasonably match the observed water contents at all depths. The mechanistic model implemented in HYDRUS-1D was the more accurate with root mean-square error (RMSE) values for water content based on all data ~0.02 cm3 cm–3, whereas for the compartment model the RMSE was ~0.03 cm3 cm–3. The results further illustrated that for a mildly heterogeneous soil (in terms of coefficient of variation for estimated hydraulic properties between soil horizons), the five-layer soil profile could be replaced by a single set of effective hydraulic properties with only a 35% reduction in performance compared with the five-layer mechanistic model. A functional evaluation of model performance using the cumulative annual drainage revealed overall good performance of the simplified models; drainage values calculated with the five-layer compartment model and the one- and two-layer mechanistic model were never more than 36% larger than their reference value. Global inverse parameter optimisation routines such as the genetic algorithm applied here are powerful tools to determine field-scale hydraulic properties of heterogeneous soil profiles for simple and complex models; model and parameter complexity can be customised depending on data availability and computational constraints.