dlawrenncar
commented
2 years ago Response from Gab: "In your context, I think the biggest benefit of using something like CLASS (or indeed the pre-closure equivalents) would be the observationally constrained uncertainty estimates that might somehow get included in the analyses in the end. Unlike a spread across competing gridded products, these uncertainty bounds represent an observationally constrained estimate. They reflect the discrepancy between in-situ measurements and the product (where observations do exist) and manage to use those observations to create a spatiotemporally complete uncertainty estimate. In reality it’s a very generous uncertainty estimate, because in addition to accounting for observed vs gridded product mismatches, it also implicitly includes the mismatch that comes from the different spatial scale of in-situ (~1km2) versus gridded data (and so site region heterogeneity plays a role). Best thought of as the expected agreement of an unseen in-situ (think flux tower) measurement within the grid cell of the final product and the grid cell value. Hopefully that makes sense."
And response from Sanaa Hobeichi @.***) who is the person who actually developed the products.
- Gab is right. I’ve looked before at the biases of the datasets that were involved in deriving Pre-DAT Rn, i.e. the pre-closure equivalent of CLASS-Rn, and they all have large positive biases against in-situ observations, these are CERES-EBAF, ERAI, MERRA-2 and GLDAS-Noah. Both Pre-DAT and CLASS have inherited some of these biases. The mean bias plot (from the CLASS paper https://journals.ametsoc.org/view/journals/clim/33/5/jcli-d-19-0036.1.xml) shows the distribution of bias across 164 flux tower sites, both positive and negative biases in CLASS are equally distributed across sites (median ~ 0), and the magnitude of positive bias is larger than that of negative bias.
[image: Chart, box and whisker chart Description automatically generated]
- I think that the benefit of comparing with DOLCE V2 https://researchdata.edu.au/derived-optimal-linear-dolce-v21/1463675/ DOLCE V3 https://researchdata.edu.au/derived-optimal-linear-dolce-v30/1697055 and LORA is that you get to compare the datasets over a longer time period, i.e. 29 years and 23 years in DOLCE V2/V3 and LORA respectively. My understanding is that CMIP6 models are out of phase, which means that years 2000-2009 in CMIP6 and CLASS are not necessarily equivalent, however, over a longer time period the comparison can be more meaningful.
On Thu, Jun 9, 2022 at 8:39 AM nocollier @.***> wrote:
Collecting discussion points for next meeting:
- Any response from Gab about CLASS low net radiation? Should we drop DOLCE and LORA? @dlawrenncar https://github.com/dlawrenncar
- Last meeting, we had some unresolved discussion about the surface soil moisture from WangMao. In particular, CESM2 and NorESM are quite wet in high latitudes relative to the data product and we wondered if this effect is real, especially given the caution provided in Dirmeyer2016 https://journals.ametsoc.org/view/journals/hydr/17/4/jhm-d-15-0196_1.xml to combine soil moisture data carefully. Here is the dataset publication Wang2021 https://essd.copernicus.org/articles/13/4385/2021/. @jiafumao https://github.com/jiafumao
- For the higher dimensional soil moisture product, @ypwong22 https://github.com/ypwong22 is working on adding models. Any updates?
- I have an initial work on an adaptation of Umakant's soil carbon data. He has several layers of data, but two I think are particularly useful to us: cSoilAbove1m https://www.climatemodeling.org/~nate/ILAMB-Test/comparisons/cSoilAbove1m.png and cSoil https://www.climatemodeling.org/~nate/ILAMB-Test/comparisons/cSoil.png which is the soil carbon above 3m. Do we compare separately to both? Umakant's estimates are high relative to the other products we have and even correlate badly with respect to the NCSCDV22. How could we check that my coarsening strategy is reasonable?
— Reply to this email directly, view it on GitHub https://github.com/rubisco-sfa/ILAMB-Data/issues/28, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFABYVCQ6FMIVZDJX57QMULVOH6YLANCNFSM5YKKQLLA . You are receiving this because you were mentioned.Message ID: @.***>
nocollier
mmu2019
ckoven
Global.Carbonbiomass product. The reference we provide discusses tropical biomass, but this is a global dataset. If you see here though,Global.CarbonandTropicalare different. I thought we had dropped this dataset? We need to restore GEOCARBON and Saatchi's Global dataset now has a reference.score = | 1 - error / bad_biome_error |, clipped to be on [0,1]. I have a comparison of the new vs old scores if the quantile used to define thebad_biome_erroris changed. I also produced a CMIP5v6 comparison. The 98th quantile is a defensible choice, but leads to scores which use very little of the [0,1] range. For that reason it seems better to use a lower quantile, but which one and what principle do we use to justify the choice?