search

VSainteuf / pastis-benchmark

MIT License
190 stars 38 forks source link

The Co ordinate reference system(CRS) for the labels and image patches #21

Closed venkatesh-thiru closed 1 year ago

venkatesh-thiru commented 1 year ago

Thank you for open-sourcing the dataset. I am currently working on a deep learning based harmonization/super-resolution pipeline for Landsat-8 and Sentinel-2 images. I would like to use the annotations from PASTIS dataset to benchmark the performance of the segmentation models when i use a combined time series of Landsat-8 and Sentinel-2 images. It would be very helpful if you can provide me with the information on the Co Ordinate Reference System(CRS) of the annotations and the image patches or some scripts which you used to prepare the dataset. The CRS information is very critical for my work, since it also involves spatial co-registration of the images between both the sources.

Also, the Sentinel-2 patches in the dataset, are they the DN values?

VSainteuf commented 1 year ago

Hi @v3nkyc0d3z, The CRS info is in the metadata.geojson file of the dataset.

What do you mean by DN values ?

venkatesh-thiru commented 1 year ago

By DN i meant the Digital Numbers, which represents the pixel values of raw Sentinel-2 raster files. In general the DN values are converted to Bottom of the Atmosphere reflectances(BOA) for L2A images. So the sentinel-2 patches in the numpy arrays, are those the DN values from the raw rasters or were there some kind of normalization before stacking the images into numpy arrays.

VSainteuf commented 1 year ago

Ok the pixel values of the numpy files are in the default range of the L2A product (no normalisation).