Possible Duplicate:
How do you decide what interpolation method to use for resampling raster data?

I want to interpolate a certiain target variable over a large territory. I trained a random forest using a sample of points on that territory where the target is known (I have many regressor layers / variables). The information of these regressors is available over the whole territory (thus the capability to interpolate the target). The thing is that the regressor layers are available in different resolutions, I used resample to force them to the same resolution (1km^2). Overall the model works rather well (This on a 20% test set). And the output map looks good. But I wanted to know if the different techiniques available for doing resampling have anything to do with the quality of the final output map? What are the pros and cons of the different resampling techiniques available in ArcGis (Nearest neighbor assignment, Bilinear interpolation, Cubic convolution). I'm only referring to continuous variables.

From ArcGis Help: Bilinear interpolation or cubic convolution should not be used on categorical data since the categories will not be maintained in the output raster dataset. However, all three techniques can be applied to continuous data, with nearest neighbor producing a blocky output, bilinear interpolation producing smoother results, and cubic convolution producing the sharpest.

marked as duplicate by whuber Jan 12 '13 at 17:38

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • It strongly depends on what do you need the data for. As usually when you build a model, your output depends on the input :) The best solution to your problem is to learn algorithms behind the the tools and decide if it is useful in your case (those 3 are pretty general algoritms and the description should be easily accesible on the internet). Of course people may help you with pointing out, how do they use those algoritms, but it does not mean that in your case it would be appropriate. Smooth or sharp properties of the output should probably not be the clues whether algoritm is right or not. – Tomek Jan 8 '13 at 19:58
  • Also I do not think that there are pros or cons for those algorithms (maybe except for the run time if you do not care to much for the accuracy of the output), all will be suitable under certain (different) conditions and requirments. – Tomek Jan 8 '13 at 20:04
  • What I was thinking is that If i have variables that have sharp edges I should use cubic. If I have a more smooth regressor I should resample using bilinear. So I was thinking of using a different resampling method depending of the type of variable. Is this worth it? What other resampling methods are there? are they more specific considering what I just said (the nature of the variable you want to resample) ? – JEquihua Jan 8 '13 at 21:09
  • You can find some useful answers in this question – nadya Jan 8 '13 at 22:51
  • Sorry, did not satisfy me fully. But thank you. – JEquihua Jan 10 '13 at 15:57