[Feature Request]: enhance accuracy of min/max statistics

Provide a clear and concise description of what the problem is.

When fetching raster data for statistical analysis, we submit a WCS GetCoverage request with a height and width given. Limiting the size of the generated GeoTIFF helps protect the server from running out of memory and also ensures a timely response. However, this can lead to inaccuracies when calculating the extreme values (min or max). WCS servers (e.g. GeoServer and Open Data Cube) generally use nearest-neighbor interpolation in order to generate the resampled tiff. If the pixel with the highest or lowest value happens not to be in in the sampled pixels, the the statistical result will be in accurate. The degree of this inaccuracy will depend on how many pixels have these extreme values. For example, weather values are generally distributed smoothly and continuously (without large abrupt changes) and therefore are more resistant to these issues. However, values that depend on landcover could experience larger changes between neighboring pixels and are more susceptible to issues with nearest resampling.

Provide a clear and concise description of what you want to happen.

Action items could include:

determining the level of effort required to add min/max resampling/interpolation to the WCS services that we use (this might require upstream contributions)
assessing adding a configuration option, so we can change the resolution of the requested GeoTIFF depending on the type of calculation
assessing pros/cons for creating lower resolution products that are generated using different resampling algorithms. For example, we could pre-generate a lower-resolution Cloud-Optimized GeoTIFF using gdal with min resampling. We would then upload this GeoTIFF to our WCS service and use that one for min calculations.

Is there anything else you can add about the proposal? You might want to link to related issues here, if you haven't already.

No response

WFP-VAM / prism-app