Impact of land use on water resources via a Gaussian process emulator with dimension reduction

The replacement of models by emulators is becoming a frequent approach in environmental science due to the reduction of computational time, and different approaches exist in the water resources modelling literature. In this work, an emulator to mimic a hydrological model at catchment scale is proposed, taking into account the effect of land use on the hydrological processes involved in water balance. The proposed approach is novel for its combination of techniques. The dimension of the temporal model output is reduced via principal component analysis, and this reduced output is approximated using Gaussian process emulators built on a conditioned Latin hypercube design to reflect constrained land use inputs. Uncertainty from both the model approximation and the dimension reduction is propagated back to the space of the original output. The emulator has been applied to simulate river flow in a rural river basin located in south west England, the Frome at East Stoke Total, but the methodology is general. Results showed that the use of the emulator for water resources assessment at catchment scale is an effective approach, providing accurate estimates of the model output as a function of land use inputs, for a reduction of the computational burden.