Questions on hydrologic model calibration methodologies

Questions around hydrologic calibration methodologies:

1) The performance of numerous rainfall-runoff models have been tested in many catchments across Australia in recent work (Petheram etal, 2009; 2012; Viney etal, 2009; Waters etal, 2013, in-prep). Two models consistently outperform the others: Sacramento (Burnash etal, 1973) and GR4J (Perrin etal., 2002; 2003). The Sacramento model tends to perform better when assessed over the calibration period, however the GR4J model has fewer parameters and the key question remains: does the simpler model allow more robust calibrations and perform better outside the calibration period?

As a side note, I am not sure whether GR4J has been tested on very large catchments (such as some we have in Northern and Western QLD). It may be that it lacks the complexity in its lower-zone storages required to model such catchments accurately.

2) Queensland hydrology has historically preferred to calibrate hydrologic reaches (as delimited by gauging stations or storages) in isolation. By using gauged flows at both the upstream and downstream ends of the reach, one obtains a good information about the routing behaviour, and estimates of historic residual fluxes (for calibration residual rainfall-runoff models, and reach losses). The benefit is that you are only calibrating to the best quality data. The downside is that the calibration period is limited to that during which data exists for upstream and downstream stations.

An alternative method is to calibrate all the models reaches together. This method introduces an amount of interplay between the parameters of successive reaches. It does mean you can use all the streamflow data (not just periods where upstream and downstream gauges have overlapping records). But it means that you are trying to calibrate many hundreds of parameters at once. It also means that you have to be careful about how to weight the relative importance of different gauges. And it means that spurious rainfall data in one subcatchment will not have an isolated effect; the resulting flows will hinder the calibration of all downstream reaches.

Which method is faster? Which is more robust (inside and outside the calibration period)?

3) The PEST optimization package has at least two optimization algorithms. The standard algorithm is very powerful, but assumes that the objective function is a well-behaved function of the model parameters. An alternative algorithm is the shuffled-complex-evolution algorithm implemented in SCEUA_P. It is slower to optimize, but makes no assumptions about the shape of the objective function, and is less likely to be trapped in local minima. A rainfall-runoff model is certainly better behaved than an operations model, but it does involve a few if-then-else blocks. So….

What optimization algorithm works best? Is the answer the same for Sacramento and GR4J?

Additional reading:

Marshall, L. (2005) “Bayesian analysis of rainfall-runoff models: Insights to parameter estimation, model comparison and hierarchical model development”, PhD Thesis, UNSW. http://www.google.com.au/url?sa=t&rct=j&q=&esrc=s&source=web&cd=5&ved=0CFwQFjAE&url=http%3A%2F%2Funsworks.unsw.edu.au%2Ffapi%2Fdatastream%2Funsworks%3A1669%2FSOURCE02&ei=XlBmUbeGK6npiAeuz4GQDA&usg=AFQjCNH27GUNdGMymyUpNi2uZfF42ihxWA&sig2=9FK49jhTAuEm35L7TwY__A&bvm=bv.45107431,d.aGc&cad=rja

Vrugt, etal. (2008)