1

I am doing some species distribution modeling with MODIS 250m data, and would like to downscale 30 arc-second (1km) bioclim data to 250m so that I can use them in my models.

My idea is this:

Rather than resampling the 1km data to 250m (e.g. bilinear interp), what if I were to:

  1. Convert the 250m grid to points
  2. Extract each of the bioclim values at each of the points
  3. Convert the points back into 250m grids.

The idea being that this is a more conservative approach than interpolation, since I wouldn't be making up new data values. Does this seem appropriate/viable, or would I be better off just resampling?

Follow up question: If I were to do this, would the resultant 250m bioclim grids still be considered continuous data? Or are they closer to discrete? The SDMs can handle either, but I need to specify for each variable.

laser
  • 966
  • 7
  • 13
cms
  • 61
  • 3
  • Why not to resample with nearest neighbor method? You will preserve data at the same way as you want without doing point / raster convertions – aldo_tapia Oct 06 '17 at 10:33
  • Thank you, if nearest neighbor is equivalent to the process I described, than I will go with it. Would you still consider the result continuous data? – cms Oct 06 '17 at 10:48
  • Is still continuous, you're only changing resolution to match with MODIS data. The aim to use nearest neighbor into continuous data is to preserve values and don't create a new dataset based on an interpolation, many researchers use this approach. But, for an altitude raster for example, you should use bilinear interpolation – aldo_tapia Oct 06 '17 at 11:17
  • Thanks for the clarification. Can you explain why you would recommend using bilinear for altitude and not for bioclim variables, even though they are both continuous? – cms Oct 06 '17 at 11:20
  • 2
    check this answer. Actually bilinear or cubic are for continuous data. The behavior os altitude is well known, but temperature or precipitation could depend of altitude, vegetation, orography, micro-climate, etc. Those data depend of too many factors, so a downscale of that kind of data could be achieved with co-kriging or other complex resampling methods, but not with a simple bilinear interpolation. – aldo_tapia Oct 06 '17 at 11:37
  • A bit too late my answer but that's fine. Assuming your covariates (250m) are highly correlated (>70% - 75%) with your response variable (1km), then you can use area-to-point regression Kriging (ATPRK) for downscaling (and not co-Kriging, as it has a lot of limitations and requires a lot of computing power and memory). Here is a link on how you can perform ATPRK in R using raster data (https://stackoverflow.com/a/73594899/13087048) – Nikos Mar 03 '23 at 12:48

0 Answers0