Remote sensing examples

[AmmarMian]

A basic example of using pyRiemann to cluster both Hyperspectral images (SPD matrices) or PolSAR images (HPD matrices).
Any input about the coding style is welcomed.

I had to rebase to get all the changes from pyRiemann master branch so I have lots of commits. Please let me know if you'd prefer I make a cleaner version starting from current pyRiemann.

[AmmarMian]

@qbarthelemy Can you check now ?
Normally I resolved flake8 errors, added your suggestions and references.

I also split SAR and hyper spectral in two files because they are two types of images. Next I'll add an example of change detection rather than clustering so that will also be a new file.

Ammar

[qbarthelemy]

Thank you @AmmarMian for all this work!
I made some small modifs, so don't forget to pull before committing and pushing.

My concern is the building time of the documentation: it went from 20 / 30 min to almost 1hour
(see https://github.com/pyRiemann/pyRiemann/actions/workflows/deploy_ghpages.yml).
Is it possible to process only a sub-image of the SAR image?

Moreover, currently, you apply a raw downsampling on images, data = data[::7, ::7], but it generates spectral aliasing.
It is better to apply a decimation, ie an anti-aliasing low-pass filtering followed by the downsampling.
We could use resize function from Pillow, but it adds a dependencies to docs requirement. Ok for you @agramfort ?

[agramfort]

I would say it's ok to add Pillow to the doc build reqs

…

Message ID: ***@***.***>

[AmmarMian]

I had tuned the downsampling so that it doesn't take too long on my PC, but I guess the ressources provided by GitHub actions are quite limited. I reduced the sampling and the max_iter a bit more and removed some methods (we don't need to compare all estimation strategies, rather I let a variable estimator that the user can change to play with that aspect).

Finally,I tweaked the scripts a bit to visualise the images from the data without downsampling and the results with downsampling. I do not really want to use filters because the data is complex-valued for SAR so it's not evident how to approach it. And more importantly, the method relies on a statistical independence of the pixels which could be compromised.

To visualize results, I propose to use pcolormeshover a spatial meshgrid since we know the resolution on the pixels thanks to the documentation of the datasets. We can then see the results over the same area and also see the lost edge because of the sliding windows approach.

[qbarthelemy]

Thanks for all the great advice about radar image processing!
Building time of documentation is now OK.

[antoinecollas]

Well done @AmmarMian 👍🏻 ! I really like that you can do everything with a sklearn Pipeline. It opens many perspectives: better evaluation, comparison of estimators, ...

Overall, you should replace pca_image by sklearn.decomposition.PCA and put it directly in PCAImage.

[AmmarMian]

Ok so from the input of @antoinecollas, I removed pca_imageand used sklearn.decomposition.PCA with a reshape. I also added the doctoring for transform of LabelsToImage.

[AmmarMian]

Thanks for the changes @qbarthelemy. Let me know if anything more is needed before merging!

[qbarthelemy]

I made some simplifications on PCA.

My last question is about RemoveMeanImage: this mean removal seems redundant with the centering applied into PCAImage.
If so, we could delete RemoveMeanImage.

[AmmarMian]

Indeed you are quite right, I used stuff I had lying around but in this case no need indeed ! We can remove this from the pipeline. Let me do that now and commit again.

[qbarthelemy]

Once again, a huge thank you for this major contribution to pyRiemann!