Errors projecting and resampling SRTM data

[TiffanyBecker]

I have SRTM data in 90m (3 arc second) resolution that I want in 1kmx1km resolution.  When I try to project the data from WGS_1984 to a relevant projected coordinate system for the area in question (Lambert Azimuthal Equal Area) AND set the output cell size to 1km at the same time, I get very strange results.  Max elevation is 30,723, but I know it should be about 4,780.  And it's not just the one strange cell over 4,780 - there are many.

The max/ min elevations look more reasonable (as far as I can tell) if I project the file, and then resample the cell size on the new file as a 2-step process.  I also get reasonable results when I project with an output cell size of 100m in one step.  So why doesn't it work as a 1-step process for 1km output?

I thought the problem might be the NODATA value of 32767.  But even when I change NODATA to 9999, I still get the same abnormally high max elevation (30,723).

I'm doing everything with bilinear resampling, by the way.

[ShaunWalbridge]

Can you share a sample of the data, or a location to download the relevant tiles? As a first start, do each analysis step separately: reproject, then resample to the lower resolution. This'll help determine what step is introducing the erroneous values.

[TiffanyBecker]

Hi,
thanks for your reply.  I'm not sure I can share the base file, it's over 500MB.  It's just mosaiced SRTM tiles, no other processing.

I did two actions separately, as you suggest.  That seems to give more or less reasonable results at both stages, after projecting and after resampling to change the cell size to 1km.  The maximum elevation dropped slightly after resampling (4767m to 4568m).

So the problem seems to happen by doing the actions simultaneously.

I could just stop here and simply use the file generated after doing the two steps consecutively (4568m max elevation version).  But I need to do the same process 19 more times for Bioclim files, to project and resample to 1km cells.   Which would mean extra work to do the two-step process.  And more importanly, if I can't figure out why I get such weird results on the SRTM file, I won't be comfortable using the revised Bioclim layers because they might also be wrong.

Any other suggestions?

[ShaunWalbridge]

If you let me know the source of the SRTM tiles, I can create a similar mosaic here to try with, or try selecting just the tiles around the erroneous high values -- I don't need the whole set, just an example area which exhibits the behavior, if you'd like to figure out the specific problem in doing both steps simultaneously.The drop in elevation you are seeing is expected: when you perform cubic re-sampling, you're averaging over a neighborhood of elevation values for each output cell.

In both cases -- Rob Hijman's Bioclim, and SRTM -- the original data providers share versions sampled at other resolutions (e.g. SRTM30). It may be possible to side-step these issues entirely by using one of these versions as your first step, as both sources have already assessed the resampling in the methods of their dataset production.

[TiffanyBecker]

Hi again,

A belated thanks for your answer.

If you want to try to replicate the steps, here's where I downloaded the SRTM tiles.  I'm using tiles 37-03 to 40-03 and 37-04 to 40-04.
http://srtm.csi.cgiar.org/SELECTION/inputCoord.asp

I mosaic'ed all tiles using Mosaic command and 37-03 as the base file.  I did not try Mosaic to New Raster, it seemed to work fine to mosaic them to an existing tile.

Then I did the simultaneous projection and resampling to 1km, which gives the weird results.

I couldn't find the SRTM30 (1 km) dataset on-line, only the metadata and reference documents.  My two step approach (project, then resample to 1km) gives basically the same max and min results as the SRTM30 dataset, just a few meters different.  So that gives me confidence that the two step process works.

You said in your post that it's normal that the max elevation would drop slightly using cubic resampling.  In fact, I am using bilinear resampling.  Would you still expect the same results after bilinear resampling to 1km (drop in max elevation from 4767 to 4568)?

Thanks,
Tiffany