A hierarchical Bayesian approach for multi‐site optimization of a satellite‐based evapotranspiration model

Abstract

Modelling is an important tool in simulating and partitioning evapotranspiration (ET). To obtain realistic partitioning of ET, a hierarchical Bayesian (HB) method was used to fit the Priestly–Taylor Jet Propulsion Laboratory (PT‐JPL) model against the multi‐tower Flux Network (FLUXNET) datasets. Unique to the HB method is its ability to exchange of information between sites and simultaneously estimate the species‐and PFT‐level parameters. The results suggested that the sensitive parameters varied at the both species and PFT levels. The parameter β (water control of soil evaporation) exhibited relatively wide species‐and PFT‐level posterior distributions, indicating that the original parameterization of soil moisture constraint may be problematic. Generally, the model with parameters determined by the HB approach showed better performance in predicting and partitioning ET than the original model, especially in evergreen needleleaf forests, open shrublands, closed shrublands, and woody savannas. To overcome the problem of parameter uncertainty (equifinality), direct observations of different components of ET are urgently needed in future studies, and assessments the extent to which the parameter uncertainties are reduced by the use of additional data.

Neglecting Model Parametric Uncertainty Can Drastically Underestimate Flood Risks

“…Bayesian calibration of hydrologic models have become increasingly popular (Hsu et al., 2009; Jeremiah et al., 2011; Kavetski et al., 2018; Raje & Krishnan, 2012; Razavi & Tolson, 2013; Shafii et al., 2015; Su et al., 2018; Zhu et al., 2018). For example, Vrugt et al.…”

“…Su et al. (2018) uses a Bayesian hierarchical model to calibrate the Priestly–Taylor Jet Propulsion Laboratory model using observed evapotranspiration measurements. Given the relatively short model run times, the hierarchical model can be fit using the Differential Evolution Markov Chain (Braak, 2006; Storn & Price, 1997), a population MCMC algorithm.…”

“…Bayesian calibration of hydrologic models have become increasingly popular (Hsu et al, 2009;Jeremiah et al, 2011;Kavetski et al, 2018;Raje & Krishnan, 2012;Razavi & Tolson, 2013;Shafii et al, 2015;Su et al, 2018;Zhu et al, 2018). For example, Vrugt et al (2008) employ an adaptive Metropolis Markov chain Monte Carlo (MCMC) sampling scheme-Differential Evolution Adaptive Metropolis (DREAM) algorithm to explore the entire parameter space of a hydrologic model.…”

“…Jeremiah et al (2011) demonstrate an improved efficiency of Sequential Monte Carlo approach over the Adaptive Metropolis MCMC samplers in exploring the parameter space where the optimal solutions lie in the tails of the prescribed prior distribution. Su et al (2018) uses a Bayesian hierarchical model to calibrate the Priestly-Taylor Jet Propulsion Laboratory model using observed evapotranspiration measurements. Given the relatively short model run times, the hierarchical model can be fit using the Differential Evolution Markov Chain (Braak, 2006;Storn & Price, 1997), a population MCMC algorithm.…”

Neglecting model parametric uncertainty can drastically underestimate flood risks

“…Bayesian calibration of hydrologic models have become increasingly popular (Hsu et al, 2009;Jeremiah et al, 2011;Kavetski et al, 2018;Raje & Krishnan, 2012;Razavi & Tolson, 2013;Shafii et al, 2015;Su et al, 2018;. For example, Vrugt et al, (2008) employ an adaptive Metropolis Markov chain Monte Carlo (MCMC) sampling scheme-Differential Evolution Adaptive Metropolis (DREAM) algorithm to explore the entire parameter space of a hydrologic model.…”