For details on changing coordinate reference systems, see Reprojection.
Up and Downsampling¶
Resampling refers to changing the cell values due to changes in the raster cell grid. This can occur during reprojection. Even if the crs is not changing, we may want to change the effective cell size of an existing dataset.
Upsampling refers to cases where we are converting to higher resolution/smaller cells. Downsampling is resampling to lower resolution/larger cellsizes.
There are three potential ways to perform up/downsampling.
If you use
reproject but keep the same CRS, you can utilize the underlying GDAL algorithms
to resample your data.
This involves coordinating the size of your output array with the
cell size in it’s associated affine transform. In other words, if you multiply the resolution
x, you need to divide the affine parameters defining the cell size by
arr = src.read() newarr = np.empty(shape=(arr.shape, # same number of bands round(arr.shape * 1.5), # 150% resolution round(arr.shape * 1.5))) # adjust the new affine transform to the 150% smaller cell size aff = src.transform newaff = Affine(aff.a / 1.5, aff.b, aff.c, aff.d, aff.e / 1.5, aff.f) reproject( arr, newarr, src_transform = aff, dst_transform = newaff, src_crs = src.crs, dst_crs = src.crs, resampling = Resampling.bilinear)
You can also use scipy.ndimage.interpolation.zoom to “zoom” with a configurable spline interpolation that differs from the resampling methods available in GDAL. This may not be appropriate for all data so check the results carefully. You must adjust the affine transform just as we did above.
from scipy.ndimage.interpolation import zoom # Increase resolution, decrease cell size by 150% # Note we only zoom on axis 1 and 2 # axis 0 (our band axis) stays fixed arr = src.read() newarr = zoom(arr, zoom=[1, 1.5, 1.5], order=3, prefilter=False) # Adjust original affine transform aff = src.transform newaff = Affine(aff.a / 1.5, aff.b, aff.c, aff.d, aff.e / 1.5, aff.f)
Use decimated reads¶
Another technique for quickly up/downsampling data is to use decimated reads.
By reading from a raster source into an
out array of a specified size, you
are effectively resampling the data to a new size.
The underlying GDALRasterIO function does not support different resampling methods. You are stuck with the default which can result in unwanted effects and data loss in some cases. We recommend using a different method unless you are upsampling by an integer factor.
As per the previous two examples, you must also adjust the affine accordingly.
Note that this method is only recommended for increasing resolution by an integer factor.
newarr = np.empty(shape=(arr.shape, # same number of bands round(arr.shape * 2), # double resolution round(arr.shape * 2))) arr.read(out=newarr) # newarr is changed in-place
When you change the raster cell grid, you must recalulate the pixel values. There is no “correct” way to do this as all methods involve some interpolation.
The current resampling methods can be found in the rasterio.enums source.
Of note, the default
Resampling.nearest method may not be suitable for continuous data. In those
Resampling.cubic are better suited.
Some specialized statistical resampling method exist, e.g.
Resampling.average, which may be
useful when certain numerical properties of the data are to be retained.