LAI calculation directly from ESA SNAP

[Yuval-Sadeh]

Feature description

@rohansaw, @wronk
I would like to ask to hear your thoughts about the way the LAI is calculated via the official ESA's SNAP platform, in compare to the existing workflow/script we are using.

Here is the link:
https://github.com/senbox-org/s2tbx/blob/master/s2tbx-biophysical/src/main/java/org/esa/s2tbx/biophysical/BiophysicalOp.java

Could you please take a look at this?
The code is open

Suggested solution

No response

Additional context

No response

[wronk]

Do they have any papers on their computation?

Also, I couldn't find the code in the repo where the computation is actually performed, but that may be just because I haven't coded in java in ~15 years.

Note that the repo doesn't seem to be maintained anymore.

[Yuval-Sadeh]

Hi @wronk, I'm not aware of a paper, but there are some documentation online somewhere.
Not sure about the code BUT! check this link out:
https://github.com/senbox-org/optical-toolbox/tree/master/opttbx-biophysical/src/main/resources/auxdata/3_0
Here they seems to have the models for LAI, including the LAI Weights Neurons in Excel files

Perhaps you or @rohansaw can find also where the actual code is or the LAI LUT they used to create the LAI Weights Neurons.

Maybe it is worth asking Jake about this, what do you think?

[wronk]

@Yuval-Sadeh, I was able to find the technical ATDB here after poking around. Page 39-47 give an overview of the architecture and evaluation.
http://step.esa.int/docs/extra/ATBD_S2ToolBox_V2.1.pdf

My sense is that we could/should more seriously consider this once the automated evaluation procedure in place (and I know you all are working on that). Using the GUI toolbox, someone could probably quickly generate some alternative LAI images from sentinel 2 data. Then we could compare how this experimental LAI estimate changes predicted yields for the fields where we have ground truth yield. Note that for any new LAI methodology, we'll probably want to think about a way to automatically optimize the manually defined hyperparameters of vercye. Some of our earlier work with Vineet should help with this.

Other notes

• The ML model itself is a pretty simple feed forward network of similar complexity as the one we're using now.
• More documentation is buried here
• Model weights and other specifics for each index look like they're specified a few layers down from the head directory here