[gdal-dev] Why is the nearest neighbour resampling method the "worst interpolation"?

[Nikos Alexandris]

On Tue, 2008-03-25 at 09:58 +0100, Vincent Schut wrote:
> William Hughes wrote:
> > Nikos Alexandris wrote:
> 
>   <snip>
> 
> > Maybe, mabe not.  The pixel values can be considered as noisy
> > samples of a continuous field.  As such, interpolated values may
> > be considered to be a better guess as to the actual value at smaller
> > pixel sizes.  On the other hand, the interpolation may have
> > unwanted interactions with your classification and/or change
> > detection methods.
> > 
> > Sometimes it helps to do your calculations with the original
> > pixel spacing, then resample for output or display.
> 
> Nikos, I tend to fully aggree with William on this. Here we also work 
> with MODIS reflectance data, and we always try to delay any 
> interpolation involving step (like reprojection etc) to the very end of 
> the processing chain. The longer you keep your pixels in the original 
> setting, the less error you possibly introduce from interpolation. We 
> usually only interpolate/reproject as a final step, e.g. before we need 
> to deliver the data and we have agreed upon a certain projection with 
> our counterpart. So we usually do not reproject reflectance data (only 
> sometimes for visual interpretation, comparison with other data, 
> presentations, etc), but only reproject interpreted data, usually land 
> cover/change classifications of some sort. For those data, NN is 
> obviously the best because you work with discrete pixel values 
> ('classes'). Also keep in mind that modis surface reflectance products 
> usually are available in 1km resolution too (and I think even lower res, 
> but I'm not sure about that), which I would prefer instead of 
> downsampling a higher res product to 1km myself. Until proven wrong, I'd 
> assume the 1km product from NASA is superior to a home-downsampled 
> 500m->1km product.

Vincent,

thank you for the details. I wish I had asked this before!

I did have to make a choice in the very beginning of my small work
(whether to reproject or not). I didn't know how to work with all MOD09
and other ancillary data in a sinusoidal projection and what would be
best: getting all data to sinusoidal, process, detect changes and go
back to a planimetric projection in order to estimate areas, etc.

I choose to reproject but with NN. I use the 250m product so in addition
I resampled the Green and Blue from 500 to 250 (using NN again). Finally
I left the Blue band out since it was too noisy. And if I guess right, I
use also SRTM3 (90m). I understand now that maybe it would be better to
interpolate the SRTM elevation data using something else than NN. I
didn't really tried that.

> On a side note and slightly off-topic: because we tend to keep our 
> spectral reflectance data in the original (sinusoidal) projection, we 
> often *do* have to reproject other misc data like elevation (srtm90). 
> Would people on this list have some knowledge or maybe hints to 
> documents on what downsampling/interpolation method would be best for 
> elevation data, for example when downsampling from 90 to 250 or 500 
> meters pixel size, keeping in mind that the result will be used to 
> calculate slope/aspect from?

> Regards,
> Vincent Schut.

I will report in case I find something useful trying out the different
methods to interpolate the SRTM3 data. But this depends also on the area
(the topography) I think.

Thank you very much.

Kind regards,

Nikos