Multi-Objective Adaptive Surrogate Modeling-Based Optimization for Distributed Environmental Models Based on Grid Sampling

Parameter optimization is needed for reliable simulations and predictions of natural processes by environmental models. The surrogate modeling-based approach is an efficient way to reduce the number of model evaluations needed for optimization. However, building a surrogate of a distributed environmental model with many output variables over a large spatial domain is computationally intensive as it involves a large number of expensive model simulations on many spatial grid cells. In this study, a novel calibration method called the multi-objective adaptive surrogate modeling-based optimization using grid sampling (MO-ASMOGS) is introduced. This method constructs the response surface surrogate of the original model more efficiently by using both parameter and spatial grid sampling. The spatial grid sampling strategy utilizes the evolutionary elitism and adaptive sampling concepts, thus allowing the surrogate model to be built using a fraction of the total grid cells over a large region. We apply MO-ASMOGS to calibrating the Noah-MP model against two surface fluxes: the gross primary production (GPP) and the latent heat flux (LH), over two plant function types (PFTs) across the continental United States. The results demonstrate that the MO-ASMOGS method can significantly improve the GPP and LH simulations. The new method needs only a small portion of the total grid cells sampled for a given PFT to achieve comparable optimization results obtained by MO-ASMO using all grid cells. This method can be very valuable in improving model calibration of computationally intensive distributed environmental models.