A fast physically-guided emulator of MATSIRO land surface model

Abstract

Current land surface models are useful tools to study the effects of global climate change on hydrological variables. However, these models tend to be relatively computationally expensive to run. This can hinder their widespread application in large ensemble modelling or uncertainty quantification studies. Therefore, it is necessary to build fast reduced-order emulators that approximate the dynamics of the original models.Here we build a fast physically based emulator that approximates the MATSIRO land surface model. The emulator equations are chosen based on review of the original MATSIRO equations, and on preliminary data analysis. The emulator uses minimal atmospheric input, and models snow water equivalent, wetland water storage, upper layer soil moisture, and total runoff on the daily time scale at 0.5° spatial resolution. Emulator parameters are optimized by fitting the emulator output to the original MATSIRO model. Parallelization allows lightning-fast computation of land surface evolution on the decadal scale. The emulator can work on the range of spatial scales, from local to regional to global. We validate simulated output with respect to the original MATSIRO model, and discuss global maps of parameter estimates. The emulator achieves excellent performance for snow water equivalent and wetland water storage. There are more challenges in simulating upper layer soil moisture and runoff, including systematic runoff underestimation over northern extratropics. We find that 25 years of training data are sufficient for reasonably fitting the emulator. The emulator can be further refined and coupled with river routing and/or economic models in future studies.