3

I know that resampling from coarse resolution to fine resolution is bad. Effectively you are making up data. However, if the study area is small are there any other options? Worldclim comes at around 900m resolution at it's finest. But running an SDM (species distribution model) on a small study area is relatively pointless as the the 900 x 900 cells are too big. The easiest option is to resample the worldclim data to 30m. Worldclim data is interpolated from weather stations. Would this count as a further interpolation or would it create nonsense outputs?

PolyGeo
  • 65,136
  • 29
  • 109
  • 338
Oliver Burdekin
  • 1,452
  • 1
  • 15
  • 27
  • What kind or resampling are you doing? If you're doing anything other than a NEAREST value, then you're probably introducing significant error ("resampling from coarse resolution to fine resolution is bad"). – Vince Mar 06 '15 at 15:43
  • 2
    @Vince Actually the opposite is true: nearest-neighbor resampling of already interpolated data is going to be one of the worst possible solutions (in terms of almost any reasonable quantifiable sense of error). For these particular data, though, it's unlikely to make any difference at all: climate variables should be changing so slowly over 900 meters that whether one resamples to a finer resolution or not is likely to have no material effect on any analysis. – whuber Mar 06 '15 at 15:46
  • I've looked at bilinear and bicubic. The area in question is very mountainous. – Oliver Burdekin Mar 06 '15 at 15:52
  • 3
    The underlying problem is that microclimates are poorly predicted by weather station interpolations. If you need such fine-resolution data you might want to consider using additional covariates to predict the data, such as insolation (derivable from a DEM), imagery, slope (also from a DEM), vegetative cover, or anything else that could be a climatic indicator and is available at finer resolutions. – whuber Mar 06 '15 at 15:56
  • 1
    If the resampled pixel size is smaller, in an even division of old pixel size, NEAREST is the only way to not change pixel values. Any other algorithm is going to smooth the edges of the old pixel boundaries. The point seems moot -- if the data is too coarse for meaningful analysis, then it shouldn't be used for analysis. – Vince Mar 06 '15 at 16:09
  • @Vince Although some data may be too coarse for analysis at a given resolution, that doesn't imply the data are altogether useless for analysis. The point is that not changing the values will generally create more error than changing them in a controlled way. The exception is when there is zero or negative spatial autocorrelation present, in which case your recommendation is optimal. It comes down to whether you believe Tobler's law applies or not--and the fact these data are already interpolated implies it does. – whuber Mar 08 '15 at 17:44

1 Answers1

2

Generally, resampling this kind of data using interpolation will probably lead to poor results, especially if the area concerned is mountainous (as whuber pointed out, microscale climate data is spatially highly variable and interpolates poorly). Increasing the resolution thirty times is however rather drastic and I'd think twice about the very relevance of such data. If you consider doing so, I'd search this answer for some methods.

However, a better way is to use cokriging or other covariate interpolation methods together with other temperature predictors - the relief most likely being the primary factor here, but others such as land cover type and insolation could be useful too.

Community
  • 1
Jan Šimbera
  • 1,314
  • 9
  • 12