This question is mainly for the validity of the comparison of the classification product of different types of spatial resolution.

Is it scientifically valid to compare very high-resolution image analysis product with that of moderate resolution. Say I have RapidEye image classification(8 class) of 2011 and forest loss pixels(just 3 class:forest loss, no forest loss, not applicable) of 2010 from landsat image products. Now I want to see what forest loss pixel of 2010 is converted to what in 2011.

Is it scientifically valid or not to compare product of different spatial resolution?


I ardently disagree with @TonyC's answer, I do not believe that your specific analytical goals are supported here. There has been a fair amount in the literature regarding comparison of landcover across different resolutions but the focus has been on a single time-period classification performance and not change across time.

The disparity between fractional cover at 5m and 30m would be significantly different and not supported in a direct comparison to quantify loss. If you were able to put a confidence interval around the loss it would likely encompass a standard deviation, as a function of uncertainty introduced by the differences resolution, and be statistically insignificant. You may be able to employ some fuzzy methods using a sampling approach and quantify the uncertainty through a Monte Carlo but, short of that, you run into a well known change of support problem called the Modifiable Areal Unit Problem (MAUP). I would also point out that you indicated that the classification schema are different as well. This is an additional issue that complicates matters and also invalidates direct comparison.

  • I think I simplified my answer too much so allow me to clarify. I agree with you 100% insofar as the caveat you pointed out. For someone who want to do such a comparison in a scientifically rigorous way, some modifications and testing are definitely needed. However, that doesn’t mean it is not worth trying at all. In some cases, especially when the data are limited, one may have to be creative and make better use of the data they have at hand. As a matter of fact, I don’t think anyone would prefer this way if he/she has datasets that support the direct comparison.
    – TonyC
    Nov 8 '16 at 18:54
  • @Jeffrey Thanks, I can compromise the classification scheme by generalization of the detailed one whereas both schemes match. Could you please clarify "encompass a standard deviation".What I have understood that the area, say agriculture,measured using 5m image will be outside of the area, agriculture, classified from the 30m by some confidence level e.g.1.96 at 95%CL. Am I right? Another one could you guide me on how to explain uncertainty using monte carlo in this case, guide some basic paper at least.
    – SIslam
    Jan 22 '17 at 13:24

Yes, it is scientifically valid, but you'll need to reconcile the projection, spatial resolution and classification scheme of the datasets that you want to compare. Also, it would be better if you can provide the classification accuracy of both maps in the paper alongside your comparison.

  • Could you just put 1/2 link of one paper pulished on IF greater than 2.
    – SIslam
    Nov 8 '16 at 16:23

I agree with @JeffreyEvans answer.

A simpler way to deal with it might be to resample rapid eye image to 30 meter and use exactly same classification scheme that is used for Landsat. Then, the comparison of the results will be valid. Make sure, when you resample, the resampled pixels geomatrically matche the Landsat pixels.

  • 2
    Your approach seems logical but you will still introduce very serious aggregation issues that would notably effect the statistical significance. Keep in mind that the goal is to quantify the loss of forest cover using different resolutions and classification schema. These differences, even if reclassified and resampled to a common raster overlay, come into play in representing the fractional cover and provide an insurmountable bias. If they were binary (forest/nonforest) classifications it would simplify things considerably and a Monte Carlo could be applied to account for uncertainty. Nov 8 '16 at 17:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.