Doubts about EOPatches for Slovenia 2019 datasets

1

Hi!
My current job is land cover classification from sentinel-2 images, and I want to use EOPatches for Slovenia 2019 datasets. But I have some doubt about the LULC label. In the blog ,the sentinel-2 images are in 2017,but in the AWS S3 Cloud Object Storage the images are in 2019, do they use the same groud truth? For the ground truth, I only find 2017 Slovenia land use map in the official websit, so if the dataset use 2019 sentinel-2 images and 2017 land use map, will it cause problems? And I can’t find the useful link of the EOPatches for Slovenia 2017 datasets.
Could you generously explain it? Appreciate your help.Thanks and regards.

Zhao Linying

2

Hi @linny1zhao!

Thanks for the inquiry. Indeed the blog posts mention 2017 data. In between we updated the example notebooks and the data, but not the blog posts, however, the principle should be the same.

But I have some doubt about the LULC label. In the blog , the sentinel-2 images are in 2017, but in the AWS S3 Cloud Object Storage the images are in 2019, do they use the same ground truth? For the ground truth, I only find 2017 Slovenia land use map in the official website, so if the dataset use 2019 sentinel-2 images and 2017 land use map, will it cause problems?

Regarding the land-cover reference: there is no problem if you use the 2017 data as we don’t expect large changes anyway. We also use the same reference in the latest notebook examples, so this should be fine.

On the website http://eo-learn.sentinel-hub.com/ you can find the land_use_10class_reference_slovenia.gpkg reference geopackage file, which contains the whole 10 class land cover reference for Slovenia which you should use.

And I can’t find the useful link of the EOPatches for Slovenia 2017 datasets.

Regarding the dataset, the format was also changed a little bit. It is not possible to download the zip file of all the 2019 eopatches, because it’s simply too much data. Instead, the eopatches were uploaded to the bucket (i.e. check example for one patch) and the bucket was made public. So it is possible for you to directly load an eopatch in your jupyter notebook from the S3 link like in the example below:

EOPatch.load('s3://eo-learn.sentinel-hub.com/eopatches_slovenia_2019/eopatch_id_100_col_7_row_7/')

So all the data should be available via a simple for loop, or something more advanced if you prefer.

Hope this helps, otherwise, let me know of any other issues that you might be having!

Cheers,
Matic

3

Hi Matic!
Thanks for your reply. This is the case that I have read some papers about land cover classification, in this papers, researchers will always use the labels which is correspond to the remote sensing images(that means the remote sensing images and the label are mapped in the same year). And the EOPatches for Slovenia 2017 datasets which I mentioned is the old eopatch datasets your team used published.Does the all dataset has been deleted.

4

The old dataset has been deleted, unfortunately. It is true that in the papers you use the labels from the same year as the reference, but this is just the ideal scenario since this is not always the case. Additionally, the change that is expected to happen in the span of a few years for a relatively developed country will not significantly change your results, so I think you are safe continuing with the available and free data.

Regards,
Matic

5

Hi Matic!
Thanks for your reply. I finally decide to do experiment on the EOPatches for Slovenia 2019 datasets. I
want to know if the datasets have been scene filtering and took cloud mask into account? And I found you have apply TFCN model on the datasets. Does the code have been publiced? If not, can you provided it for me?
Appreciate your help.Thanks and regards.

Zhao Linying

6

Hi @linny1zhao

You can find the code for the TFCN in eo-flow, with an example usage.

The input images are interpolated frames considering the cloud masks, similar to what we did in the land cover example.