Biophysical Processor Expression of Concern with algorithms

Optical Toolbox
1

Hello. I am writing to advise users of the S2 Biophysical Processor about two concerns with the algorithms used. I realize that this is not a ‘production’ issue but it a comment about data quality.

Richard Fernandes
Canada Centre for Remote Sensing

There are two issues.

First, the processor does not exactly implement the Simplified Level 2 Prototype Processor algorithm described in Weiss and Baret, 2016 (https://step.esa.int/docs/extra/ATBD_S2ToolBox_L2B_V1.1.pdf) due to an error in how the calibration database was produced. This was verified by Dr. Marie Weiss. It will result in overestimates of LAI, fAPAR and fCOVER at high (>4) LAI as has been observed during validation.

We have implemented an exact version of the SL2P in the LEAF-Toolbox on Google Earth Engine (GitHub - rfernand387/LEAF-Toolbox). It does not exhibit a bias over non-forests and, as expected to to lack of clumping, underestimates LAI over forests. Please consider verifying any L2B products from the biophysical processor using LEAF.

Second, I have noted there is an option for applying the biophysical processor to 10m resolution bands only. Validation of 10m products is sparse but indicates uncertainty does not satisfy Copernicus Global Land Service requirements. I would encourage users of the 10m processor to conduct validation over their regions of interest since our evaluation of the SL2P algorithm with only 10m bands suggests it is unlikely to meet these requirements without additional developments.

1 Like
2

Hello,

Thank you for feedback.
You refer to the old ATBD implementation v1.1.
Is it concerned the version v2.1 too (http://step.esa.int/docs/extra/ATBD_S2ToolBox_V2.1.pdf)?

3

Dear Florian,

These issues are a concern for V2.1. The issue was due to a bug in the original MATLAB code from Marie Weiss that incorrectly specified simulations used to calibrate the retrieval algorithm. To the knowledge of both myself and Marie this bug was never addresed in the Biophysical Processor. I have verified it persists since there are differences in a recent (V2.1) output from SNAP and my code which implemnents the correct SL2P algorithm that we (including Marie) have used for validation Validation of the Sentinel Simplified Level 2 Product Prototype Processor (SL2P) for mapping cropland biophysical variables using Sentinel-2/MSI and Landsat-8/OLI data - ScienceDirect
Validation of baseline and modified Sentinel-2 Level 2 Prototype Processor leaf area index retrievals over the United States - ScienceDirect

Note this issue was presented at the Sentinel 2 4th validation Team Meeting in 2021 and we will present if again in the 5th meeting in April.

As for the 10m resolution product. I expect the bug is also there BUT more importantly the issue is that there has been no published validation of it that indicates is is reasonable. I suggest ESA support processors that have reasonable validation or at least indicate it is not validated.

Regards

Richard

4

Hi,
Firstly, thanks to the SNAP team for releasing SNAP 9 :slightly_smiling_face: I noticed a couple of items related to the biophysical processor in the Sentinel-2 Toolbox change log but nothing relating directly to this issue. Was it fixed in SNAP 9 or is it still being investigated?

5

Hi,
Your link concerns the S2 cache deletion SIITBX-288.
There is no improvement for the biophysical since 8.x version.
A new version algorithm of the INRAE should be provided soon.

6

Thanks for the reply. I meant to link to the full list of issues resolved in SNAP 9 (which can still be seen using the above link), and not the specific cache issue. At least three of those have “biophysical processor” in the name. In any case, I am looking forward to the new version :slight_smile:

7

As a follow up please refer to this paper that documents the issue in more detail. The paper also provides a more general performance assessment of the “corrected” SL2P over forests. Validation of Simplified Level 2 Prototype Processor Sentinel-2 fraction of canopy cover, fraction of absorbed photosynthetically active radiation and leaf area index products over North American forests - ScienceDirect

Note that the difference between the SNAP implementation and the ATBD specification is not large for low LAI but can sometimes be large for LAI>3.

8

Thanks @rfernand387 for interesting paper! Is there a Python implementation of SL2P?

@FlorianD @mengdahl is this issue going to be addressed in SNAP 10? Looking at Figure 3 in Richard’s paper the problem is quite severe and the bug is well identified so should be easy to fix. The SNAP biophysical processor is probably widely used in research and applications and this issue has been hanging around for at least year and a half now so I hope it’s on the high priority list.

9

Dear all.

You can find the “correct” version of SL2P implemented here Home · rfernand387/LEAF-Toolbox Wiki · GitHub. It relies on the Google Earth Engine API but the neural network coefficients can be easily extract ad applied in open source code if required. Feel free to contact me if you need information.

10

@diana_harosa Please capture this issue in JIRA if it hasn’t been captured already.

1 Like
11

Hi all,

Is there any progress in this regard?
As @radosuav mentioned the biophysical processor is commonly used among scientific and business communities, thus it is very urgent to solve this issue.

Best
Paulina

1 Like
12

JIRA ticket SNAP-3610 created.

13

Thank you @diana_harosa, we´re looking forward to your updates regarding the improvement of the biophysical processor!

Cheers, Caroline

14

Thanks @diana_harosa. Could you indicate roughly by when can SNAP users expect the corrected version of SL2P, to evaluate if it is acceptable for running projects that make use of it, or if it is necessary to plan alternative solutions? Thank you

15

Dear @mariapina_castelli ,

Unfortunately, this fix will not be included in the upcoming SNAP release (10.0.0) as we are currently in the final testing stage, but it will be included in SNAP 11.