I am relatively new to working with satellite imagery at 1 m resolution and am struggling with a methodology for "grouping" the imagery.

Using ArcGIS, I have performed an ISO classification on some aerial imagery and am trying to categorize different vegetation types. I am "see" the delineations between vegetation types, but they are based on multiple categories from the ISO classification. Combining the classes does not help. I have also worked with segmentation mean shift, but this yielded less than satisfactory results.

I am wondering if these is some methodology that allows for the reclassification of a classified raster that takes into account the class of neighboring cells. I know this is a relatively open-ended question, but not sure how to clear it up.

I think that the solution would be akin to classifying an impressionist painting.

  • do you think hot spot analysis would work for you? I might not have understood the question though. I would prefer Erdas Imagine for any remote sensing/ imagery analysis. – Liverpoolaki Dec 21 '17 at 9:23
  • Maybe, but the classes in the classified image are categorical and not in an order ( I think. That is the problem with ArcGIS, figuring out what functions are actually doing!). I think the real path forward is to use the original multiband raster (RGB + NDVI + texture) and use some type of deep learning algorithm. I am trying to use the classified 1 band raster and then perform image classification on the resulting image. Does this make sense? – user44796 Dec 21 '17 at 14:24
  • Have you considered a loop over each specified class to produce multiple euclidean allocation rasters which could then be map algebra'd (summed) together? Once this is done, you would need a complex matrix of possible solutions to link back to your data in a large CON statement. Each allocation raster would have to contain unique values on to itself to facilitate the matrix. This is an ugly solution, but could work if you know your relationships between neighbors and the sources. – D_C Dec 21 '17 at 20:31
  • I am not sure this would work, as I really don't know the relationships between neighbors and sources. What I am really looking for is a way to classify the 1 band unsupervised classification. – user44796 Jan 4 '18 at 15:20

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.