Values change after re-projecting the NetCDF files to TIFF file

Question

I am trying to re-project a NetCDF file to a uniform grid EPSG: 4326 and convert it to a TIFF file. However, after the GDAL re-projection the values in the file change, i.e. the values are not retained after the re-projection takes place. The commands that I am using are:

gdal_translate HDF5:"LC08_L1TP_029037_20190611_20190611_01_RT_L2GEN_2.nc"://geophysical_data/Rrs_443 -a_nodata -32767 output.tif

gdalwarp -t_srs EPSG:4326 -srcnodata -32767 output.tif reprojected.tif

The links to the files are:

1. NetCDF file: https://drive.google.com/open?id=1XKNOqf_-MIcgR-f_qBWP-92ZNP3pEsgj

2. output.tif file: https://drive.google.com/open?id=1nMcMIRiOVyiwjc8tHegzRyHxn2K9762o

3. reprojected.tif file: https://drive.google.com/open?id=1Uh_UfWkeiEDuwKvDp4qCTjyBh_Bra6cZ

Answer 1

A raster is a set of cells that form a grid; each cell has a value. When you reproject a raster, you are re-drawing the grid to be aligned with a new projection. So, in the below figure, your original raster grid is shown in blue, and the reprojected grid is shown in red.

Right away you can see a problem--the grids do not align. So, for example, in the zoomed in portion, you can see that the black cell (which is part of the reprojected grid) actually overlaps five of the original grid cells (with values of 4, 0, 1, 2, and 1). So the question is: what value do we assign the cell with the question mark (?)?

That is where resampling comes into play. You must choose a method for combining the five contributing pixels to assing the value to ?. Default is "nearest neighbor", so the algorithm will assign the ? the value of the nearest blue cell (2 in this case). Or you could choose "average", which will simply average all the contributing blue cells to assign ?. Or linear interpolation, or any of the many options gdal_translate provides.

The bottom line is that (and someone feel free to correct me here) it is impossible to have a reprojected image that has the same values as the original in each location (otherwise it would not be a reprojection). You must resample your data to the new grid.