DA notebook

[jgomezdans]

• Model exploration with parameters ("sliders")
• Automatically fit the LAI ("DA button")
• Plot yield distribution for all fields w/boxplots
• Per pixel LAI exploration
  • Show yield
  • Show parameter spatial distribution

[jgomezdans]

Need to update GUI and test "DA button" with big safe ensemble Tuesday.

[jgomezdans]

I have tested the DA button, but my results are very different to yours @Ma-hy

Looking into your ensemble fitting code, @MarcYin noticed that you calculate the wrong LAI value:

np.nanmean(np.transpose(obs['mean_bios_all'][:,4,:])*obs['mean_bio_scales_all'][:,4], axis=1)

the LAI should be calculated as

f.f.mean_bios_all[:, 4, :].mean(axis=0)

We need to discuss this tomorrow morning 9am. I've sent an invite through.

[profLewis]

That’s not good. Please keep me informed how this goes this morning. I hope it’s just something trivial. I was thinking that if you are still having trouble, then a pragmatic solution would be to apply some tolerance to fengs empirical model for each sample as a first pass. That would ensure that all candidates broadly follow that pattern. It would be better not having to impose that, as the danger is that you are ‘just’ applying the empirical model. Which isn’t really what we want. I think it would make it difficult to write it up. If you do do that, check that the result isn’t just exactly the same as the empirical model then. Feng - can you give them the coefficients for your empirical model, just in case. Lewis Professor Lewis, NCEO/UCL

…

________________________________ From: José Gómez-Dans ***@***.***> Sent: Tuesday, March 1, 2022 7:17:09 PM To: UCL-EO/Workshop2022 ***@***.***> Cc: Subscribed ***@***.***> Subject: Re: [UCL-EO/Workshop2022] DA notebook (Issue #15) I have tested the DA button, but my results are very different to yours @Ma-hy<https://github.com/Ma-hy> [image]<https://user-images.githubusercontent.com/139304/156233269-f330f347-8882-46c8-bca1-18f17b5eb442.png> Looking into your ensemble fitting code, @MarcYin<https://github.com/MarcYin> noticed that you calculate the wrong LAI value: np.nanmean(np.transpose(obs['mean_bios_all'][:,4,:])*obs['mean_bio_scales_all'][:,4], axis=1) the LAI should be calculated as f.f.mean_bios_all[:, 4, :].mean(axis=0) We need to discuss this tomorrow morning 9am. I've sent an invite through. — Reply to this email directly, view it on GitHub<#15 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAOC3GUQUN3RSALVBQ4YMQTU5ZULLANCNFSM5PKEBJSA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[profLewis]

Also, can you do plots of the parameters? We surely need to see these to be able to interpret Lewis Professor Lewis, NCEO/UCL

…

________________________________ From: Philip Lewis ***@***.***> Sent: Wednesday, March 2, 2022 7:41:06 AM To: UCL-EO/Workshop2022 ***@***.***>; UCL-EO/Workshop2022 ***@***.***> Cc: Subscribed ***@***.***>; Feng Yin ***@***.***> Subject: Re: [UCL-EO/Workshop2022] DA notebook (Issue #15) That’s not good. Please keep me informed how this goes this morning. I hope it’s just something trivial. I was thinking that if you are still having trouble, then a pragmatic solution would be to apply some tolerance to fengs empirical model for each sample as a first pass. That would ensure that all candidates broadly follow that pattern. It would be better not having to impose that, as the danger is that you are ‘just’ applying the empirical model. Which isn’t really what we want. I think it would make it difficult to write it up. If you do do that, check that the result isn’t just exactly the same as the empirical model then. Feng - can you give them the coefficients for your empirical model, just in case. Lewis Professor Lewis, NCEO/UCL

________________________________ From: José Gómez-Dans ***@***.***> Sent: Tuesday, March 1, 2022 7:17:09 PM To: UCL-EO/Workshop2022 ***@***.***> Cc: Subscribed ***@***.***> Subject: Re: [UCL-EO/Workshop2022] DA notebook (Issue #15) I have tested the DA button, but my results are very different to yours @Ma-hy<https://github.com/Ma-hy> [image]<https://user-images.githubusercontent.com/139304/156233269-f330f347-8882-46c8-bca1-18f17b5eb442.png> Looking into your ensemble fitting code, @MarcYin<https://github.com/MarcYin> noticed that you calculate the wrong LAI value: np.nanmean(np.transpose(obs['mean_bios_all'][:,4,:])*obs['mean_bio_scales_all'][:,4], axis=1) the LAI should be calculated as f.f.mean_bios_all[:, 4, :].mean(axis=0) We need to discuss this tomorrow morning 9am. I've sent an invite through. — Reply to this email directly, view it on GitHub<#15 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAOC3GUQUN3RSALVBQ4YMQTU5ZULLANCNFSM5PKEBJSA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Ma-hy]

Tried assimilating correct LAI, like below

[Ma-hy]

Compared with the wrong LAI calculation

[jgomezdans]

Your new thing is the same as my plot above: only a small variation between 2200-2500 kg/ha. Plot them with the same axes.

[Ma-hy]

Yeah, the slope is quite flat, but keeps high R value, by the way, the pixel-by-pixel assimilation shows a bit larger variation.

[profLewis]

So what is the way forward here? It sound like you don’t have a viable DA for the maize crop. Do we have a viable workshop then? Did you try filtering from the empirical relationship? Lewis Professor Lewis, NCEO/UCL

…

________________________________ From: Hy Ma ***@***.***> Sent: Wednesday, March 2, 2022 1:59:49 PM To: UCL-EO/Workshop2022 ***@***.***> Cc: Professor Philip Lewis ***@***.***>; Comment ***@***.***> Subject: Re: [UCL-EO/Workshop2022] DA notebook (Issue #15) [image]<https://user-images.githubusercontent.com/17977287/156375772-f2b64fbd-41a2-498e-a44e-8b3521efd702.png> Yeah, the slope is quite flat, but keeps high R value, by the way, the pixel-by-pixel assimilation shows a bit larger variation. — Reply to this email directly, view it on GitHub<#15 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAOC3GSXUXSEHJLFDA23MG3U55X5LANCNFSM5PKEBJSA>. You are receiving this because you commented.Message ID: ***@***.***>

[jgomezdans]

I'm currently filtering by the empirical relationship and will have some results in a bit. the ensembles from Hongyuan didn't cover empirical relationship line, so have changed things a bit to get to that point.

[jgomezdans]

OK, so after fitting the empirical line, and changing a more sensible set of parameters, here's something more hopeful. Results first:

• Fit to the LAI is fair. Remember, small ensemble (made smaller by filtering, around 650 ensemble members)
  
• Observed vs predicted yield: Pretty good correlation, slope around 0.45
  
• Parameters vs obs yield. Reasonable trends (see for example CVO), but there's a lot of internal correlation here
  
• Correlation matrix shows huge correlations, so we may be able to further simplify the number of crop parameters
  

I'm doing this on a small ensemble, so I selected the 5 best solutions (for the correlation matrix, I selected the best 20). In general, this looks reasonable, and I think we can maybe expect better results with bigger ensembles that allow better fitting to LAI, and possibly reducing the input parameters space (not too bothered about that).