Several recent contributions to the hydrologic literature have demonstrated an inability of
standard model evaluation criteria to adequately distinguish between different parameter sets
and competing model structures, particulary when dealing with highly complex
environmental models and significant structural error. The widespread approach to model
evaluation that summarizes the mismatch, En = {ek;k = 1,-¦,n} = Yn -ËYn between n
model predictions, Yn and corresponding observations, ËYn in a single aggregated measure of
length of the residuals, F not only introduces equifinality but also complicates parameter
estimation. Here we introduce the Differential Evolution Particle Filter (DEPF) to better
reconcile models with observations. Our method uses sequential likelihood updating to
provide a recursive mapping of {e1,-¦,en}- F . As main building block DEPF uses the
DREAM adaptive MCMC scheme presented in Vrugt et al. (2008, 2009). Two illustrative
case studies using conceptual hydrologic modeling show that DEPF (1) requires far fewer
particles than conventional Sequential Monte Carlo approaches to work well in practice, (2)
maintains adequate particle diversity during all stages of filter evolution, (3) provides
important insights into the information content of discharge data and non-stationarity of
hydrologic model parameters, and (4) is embarrassingly parallel and therefore designed to
solve computationally demanding hydrologic models. Our DEPF code follows the formal
Bayesian paradigm, yet readily accommodates informal likelihood functions or signature
indices if those better represent the salient features of the data and simulation model. |