Sensitivity analysis

[iMarang]

Hi, I'm not sure this counts as an issue, but I wanted to confirm if this approach/results are valid.

I have run a sensitivity analysis on the Wofost72_WLP_FD model using the SAlib python implementation of the Sobol variance decomposition technique. I’ve run something similar over APSIMx previously.

I did this for block sizes 100, 200, 300, 500 and 1000 equating to between 5400 and 54,000 model simulations. I ran the analysis for end season production (‘TWSO’ or yield) as well as in season plant characteristics (LAI and SM on 5 dates). I want to use this analysis both for identifying parameters to calibrate/fix as well as for use in a modelling application. I’ll write the analysis up for journal publication eventually (and include yourself as co-author if you’re interested), however, right now I just would like a sanity check as to whether these results are reasonable.

Attached are plots showing the main and total effects (left hand axes) and interaction effects (right hand axes) for each parameter for each block size. I’m a little surprised by the results, however, the general pattern is consistent across the block sizes and the confidence intervals (not shown) reduce as expected over increased simulation runs (highest is 0.1). I’m particularly interested in why Specific Pod Area (SPA) is so influential for yield. I thought this was used as ancillary photosynthetic assimilate production alongside LAI?

The parameter ranges and the associated site description and weather files used to run these simulations are also attached.

WOFOST72_Sobol_plots.zip Parameter_and_forcings.zip

Let me know what you think, are these sensitivity indices reasonable?

Best, Ian

[iMarang]

Addressed offline, thanks Allard.