Processing at 10 m resolution
Processing at 10 m resolution
ACOLITE currently processes all Sentinel-2 data at 10 m resolution. The pixels from the 20 and 60 m bands are replicated without interpolation 4 and 36 times to form a 10 m grid. When no cropping is applied, the output has dimensions of 10980x10980 pixels.
Re: Processing at 10 m resolution
Hi Quinten,
Is the resampling method described in your post above still valid with new generic version of Acolite (with default settings and without specifying a 20 or 60 m target resolution) ?
I’m interested in the 10 m resolution QAA outputs obtained with Acolite when processing Sentinel-2 images, in particular because the Sentinel-2 coastal band at 443 nm is natively available at 60 m only.
Many thanks in advance !
Thibaut
Is the resampling method described in your post above still valid with new generic version of Acolite (with default settings and without specifying a 20 or 60 m target resolution) ?
I’m interested in the 10 m resolution QAA outputs obtained with Acolite when processing Sentinel-2 images, in particular because the Sentinel-2 coastal band at 443 nm is natively available at 60 m only.
Many thanks in advance !
Thibaut
Re: Processing at 10 m resolution
Hi Thibaut
This post referred to the IDL implementation, and was indeed the intended method for processing in Python.
In the new generic version this resampling is done correctly, replicating the coarser resolution data when upsampling, and performing neighbourhood averaging when downsampling. Both are performed automatically for the different resolutions using the gdal.Warp function in acolite/shared/read_band.py. In the previous Python version I was using scipy.ndimage.zoom with a 0th order spline interpolation, effectively a NN interpolation. This method introduced a 1/3 pixel offset when upsampling, presumably due to issues with pixel centre/edge interpretation. For downsampling a NN was used in the previous version, which was perhaps not ideal!
To summarise, in the generic version upsampling (60m -> 20m, 10m, 20m -> 10m) is done by replicating the pixel values, and downsampling (10m -> 20m, 60m, 20m -> 60m) is done using neighbourhood averaging.
Hope this is clear!
Quinten
This post referred to the IDL implementation, and was indeed the intended method for processing in Python.
In the new generic version this resampling is done correctly, replicating the coarser resolution data when upsampling, and performing neighbourhood averaging when downsampling. Both are performed automatically for the different resolutions using the gdal.Warp function in acolite/shared/read_band.py. In the previous Python version I was using scipy.ndimage.zoom with a 0th order spline interpolation, effectively a NN interpolation. This method introduced a 1/3 pixel offset when upsampling, presumably due to issues with pixel centre/edge interpretation. For downsampling a NN was used in the previous version, which was perhaps not ideal!
To summarise, in the generic version upsampling (60m -> 20m, 10m, 20m -> 10m) is done by replicating the pixel values, and downsampling (10m -> 20m, 60m, 20m -> 60m) is done using neighbourhood averaging.
Hope this is clear!
Quinten
Re: Processing at 10 m resolution
Hi Quinten,
Sorry for late reply, and thanks a lot for your answer, very clear !
Have a nice day,
Thibaut
Sorry for late reply, and thanks a lot for your answer, very clear !
Have a nice day,
Thibaut