Calibration strategy and updates

[khufkens]

Open questions on where to go with the calibration routine, and some of the practical decisions that need to be made there.

• selection calibration time frames
• selection calibration method (stick to GenSA or go to BayesianTools)
• ...

[imfeldn]

Further discussion points:

• sampling procedure (inside time frames)?
• think about selection of sites due to highly heterogeneous area + changes in data quality over time??
• decide on "experiments", e.g. time frames, calibration methods, are there more?
• adaptations in model evaluation

[khufkens]

Duration of calibration?

Half a weekend for 10 phenology models, 3 seeds, and 7 indices.

Todo:

Pick bayesiantools to get to the parameter uncertainties (check rsofun implementation).

Sampling scheme for stations:

• not all stations run full length, which ones do? @imfeldn
• representation across climate for the ones that are incomplete, what would the sampling scheme be to balance the datasets across decades?
• Every 2-decades sampling period? OK? or longer, 30 to 50 years or so? Too long will regress towards the mean
• Check coverage of phenology phases across stations and species
• ....

[imfeldn]

The following graph shows the observations per year if we select only series that have a coverage of at least 75% per period across the following periods: 1961-1990, 1971-2000,1981-2010,1991-2020.

Thus, from 1961 to 2020 we'd have a rather consistent set of series, depending on species/phenophase of around 30 series in total (and ~900 observations per period). It's a bit hard to make a consistent binning of series based on e.g. their climatology, just because there are very few series in the first few years from 1951 to 1960. Instead of binning by climatological mean, maybe we just can stick to the constant station/series network, still use the period 1951 to 1980 for calibration but discuss that for this period, results may deviate because of a lower number of samples.