Dear Peter

If you chose hight resolution for all images, re-sample before classif it's beter.

Otherwise (small resolution), no matter (before or after it gives the same classification results).

good luck

Dear Peter,

Better to re-sample before performing classification. if you do reverse the probability of your data lose will be very high.... and your purpose may not fulfil  due to information lose in classification image.

Cheers,

Saptarshi

You should NOT re-sample to the higher resolution of both images for change analysis because the accuracy of your result can be no better then the lowest accuracy of any input; otherwise you are committing an ecological fallacy. I would re-sample (down-sample) the higher resolution image AFTER classification. 

Hello Peter,

Re-sampling to finer resolution should not change the information content in the image, and therefore should not change the quality of the change detection. So far as your overall change-detection performance is concerned, it should not matter when you re-sample. The choice is a matter of practical convenience of implementation alone.

Note that this does not mean that you will get absolutely identical results if you re-sample before or after classification. It means that the expected (mean) value of any plausible statistical measure of the quality of the change detection will not change in a statistically significant way if you follow one option or the other. 

In order for a change-detection quality measure to change in a statistically significant way, re-sampling would have to be systematically adding (or losing) information as a rule in one option relative to the other, which should never happen with re-sampling.

Ronald

I got the impression that Peter is asking about re-sampling in order to change the cell size (map scale). Let's say that one image is at 60 m and another is at 30 m. You can't produce a change map at 30 m because you didn't have 30 m of spatial detail for both images.

Information content is scale-dependent. If re-sampling to change a projection then no information is lost; however, if re-sampling to change map scale then information can indeed be lost. Re-sampling from 30 m to 60 m does lose information. Keep in mind that re-sampling 60 m to 30 m does not gain any information, and that information would be necessary to compare to another map at native 30 m resolution. 

Resampling always involves data loss, whether for a map projection change (since different numbers of input pixels can affect the various output pixels because of geometric distortion), or a resolution change going to a coarser cell size. You just have to do it judiciously. I believe John is correct in presuming that Peter wants to resample to coarsen the data, and John is again correct that the finer scale raster should be coarsened to match the coarser raster (not the other way around).

Definitely resample sensor (i.e., satellite) data before classifying. (And if you need to reproject the data, do that before resampling.) If the data is continuous, ratio data (e.g., RGB or multi-spectral bands or DEMs are), it's probably best to use a computation-intensive method like cubic convolution, or at least bilinear interpolation (not nearest neighbor). The problem with resampling the classified raster afterward is that you're no longer able to resample the actual data, just your already abstracted, statistically-binned model of it. And if your classification is nominal (e.g., land cover classes), you can really only do it based on frequencies (e.g., majority neighbors), rather than on sensor-derived ratio data (e.g., RGB intensities). So resampling before classifying introduces less analytical error.

In case you want to resample to fine resolution, consider pansharpening with panchromatic band.

Resampling generally done before the classification. 

I agree with sabelo , you can go through pan sharpening also