I want to assess the accuracy of a settlement classification I made (classification map) by comparing it to another product's settlement classification (reference map). Both maps have two "normal" classes, but also have large areas where the land use is not classified, so they rasters are full of noData values.

For the error matrix, do I convert noData into a new class for which sample points will be created? And do you then also calculate overall, producer, user accuracy etc. for this noData class.

I didn't find any real info how to deal with noData in these assessments, but got the impression that people actually create these sample points but do not include to calculate the performance indices...

  • 1
    My gut feeling is to not sample the NODATA areas as these are areas you are not creating with your classification. If you treat NODATA as a class I suspect it will bias the final assessment. If you did a random sample across the image I guess most will end up in nodata areas, so some sort of stratified sampling is probably better? I'm no expert on this so it would be prudent for you to wait and see if people with more experience comment. – Hornbydd Jul 13 '18 at 10:17
  • Let's say you do a stratified sampling, so you don't get any sample points in your NODATA cells in the reference map, but of course you a lot of these points are in the NODATA areas of you classification map. How are these then dealt with in the error matrix? For example is the overall accuracy still calculated over the total number of points or is it a new total (old total - number of points that are NODATA in the classification map)? – Johanna Jul 13 '18 at 10:32
  • 1
    I would leave out the NoData values. In general, i would avoid creating sample points on areas in the image that are of no interest (i.e. NoData cells). The stratified sampling method you have mentioned should ONLY be used within the boundary of the cells that are of interest. As @Hornbydd mentioned, the Nodata will obfuscate your confusion/error matrix. – Jonas Jul 13 '18 at 11:13
  • But this doesn't answer the question what to do with the points that then fall in the noData areas of the classification map. There is no point in stratifying so that you only have points from the beginning that fall in "class" cells of both maps. Or am I totally misunderstanding something here? – Johanna Jul 13 '18 at 11:21
  • If you were to clip the classified map to the extent of the reference map, would there still be NoData values within the classified map? – Aaron Jul 13 '18 at 19:45

Your Answer

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

Browse other questions tagged or ask your own question.