3

The results of an unsupervised image classification may vary A LOT depending on the variables you use from the beginning. These are the variables for the tool in ArcGIS:number of classes, minimum class size & sample interval. enter image description here

But you can reply me in any homologous tool of the software of your choice.

I'm interested to know how do you decide these variables. For instance, it is said than the minimum class size should be approximately ten times the number of bands.

But what about the others?

In my experience there are not any "fixed" rules and it is more a test-error process. But I'd like to know how other users decide these variables and if you follow any special rule.

  • 3
    could you please refine your question in order to focus on one specific question. There are at least 4 subquestions (atmospheric correction, normalisation, number of classes, feature selection...) here and different people might address only one or two. I am sure that you would have received more attention if you could focus. – radouxju Sep 4 '15 at 12:55
  • I understand your point @radouxju. The thing is I'm interested in any information you can give me. If you just tell me you have no method to decide, let's say, the number of classes and you just try and try until you see a convincing result for you, this information is also useful to me. If you mention some (out of the regular) tool you found to help you in some way do an image classification technique it will be also helpful, anything will be valuable intel for me. It isn't a yes or no question or an specific one, I know it may be more difficult to answer and that's why I added a bounty. – Albert Sep 4 '15 at 13:49
  • 1
    "It isn't a yes or no question or an specific one" means to me that this is too broad for our focussed Q&A format. I agree with @radouxju that you would be more likely to get an answer (with or without a bounty) by asking a focussed question. My understanding is that I could remove this bounty and close the question as too broad (see Shog9 answer to meta.stackexchange.com/questions/121448/…) but until the bounty expires I would prefer you to be the one to address the issue of broadness. – PolyGeo Sep 4 '15 at 23:49
  • 3
    I didn’t try to “protect” anything. You paint it like if I plotted some Machiavellian plan to con you all. I just asked a question to get some answers and, later on, “paid” a bounty because I wasn’t obtaining any (by the way, question edited twice by yourself before any bounty). Is that simple. I may have been wrong in the approach of the question but there was certainly no bad intention. I’ll try to edit it and see if it can fit better in your policies. @radouxju I’ll try to be more precise, and also make sure the question is findable. Thank you all for the advices! – Albert Sep 7 '15 at 16:13
  • 1
    I would find all of this pretty interesting. I don't know anything about unsupervised classification, but I would like to learn. If potential answers in here are able to direct me to areas to focus learning, it would be pretty useful. I tend to learn best when I see something like this and there are multiple people providing different viewpoints/experiences. – Branco Sep 9 '15 at 12:54
2
+50

As noted by others, it is difficult to come up with a hard and fast rule for remote sensing analyses. There are many different factors to consider based on what your end goal is, the type of environment (desert vs jungle vs urban), complicating factors (clouds or terrain or shadow effects), and if you are conducting a temporal analysis.

In the past I looked for comprehensive, general tutorials, and never found one. What I did find is that you generally have to piece together various best practices and fit them to your use case. One option is to search for a scientific paper that has addressed a similar point, and read through their methods. They likely will not list the exact tools they used, but you may be able to reverse engineer similar methods using your software of choice.

All that being said, to try and briefly address the variables in the ArcGIS tool you list:

  • Bands: As noted by others, this depends on many factors. Choose bands or indices that will help identify your topic of interest, e.g. urban areas, vegetation health, land/water. See USGS and Landsat 8 band combinations.
  • Number of classes: Again, this depends on your project specifics. If you want a water/land mask, you only need two classes (although in practice, I might choose ~3-4, then run the tool again to aggregate areas, e.g. shadows that could be incorrectly identified as water). The quick and fast rule is: Pick the fewest classes that you need. Generally the fewer the classes, the more accurate. For example, having forest/non-forest will almost certainly be more accurate than attempting to classify evergreen, deciduous, mangrove, barren land, etc.
  • Number of Iterations: Try the default for now. Fewer iterations will run faster.
  • Minimum Class Size:: Again, this depends on your use case. In essence, you are saying what the minimum area for a class needs to be. This will depend on the area of analysis, land cover/use types, and how generalized you want the final classification to be.
  • Sample Interval: Once again, this can vary. Leave it as the default and see how the comes out. ESRI has more information on this, and classification in general here.

Sorry to not be of more assistance. Again, it may be frustrating to hear, but remote sensing is very dependent on project specifics. It looks like you have access to ArcGIS, so I would say give that a try and see how it comes out.

If you have time, you may be interested in other programs that I would argue are better suited to the task. One fairly user-friendly option is provided as as a QGIS plugin and the author put together a tutorial. Other options include Opticks and GRASS (both free) and ENVI and eCognition (both commercial). If you are familiar with ArcGIS that will be the most straightforward for now. However, if you plan to keep doing or need more advanced analysis capabilities I highly recommend exploring other options, as ArcGIS is not really intended as a fully-featured image-analysis program.

  • Thank you Dan for your answer! It is true that sometimes is frustrating seeing there's no precise or specific tutorials about this topic but it also makes it much more fun. Reading your answer I have come out with some thoughts/ideas (not necessarily good ones). One is about the band combinations. I always perform the analysis with as many bands as possible but, is it possible that some bands remove inter-variation in some classes? Would it be better to classify, let's say urban with bands 764, mask the patches resulting and then classify again? – Albert Sep 10 '15 at 7:00
  • This would serve to "stretch" the variation among the rest of the classes, you could do this process repeatedly. So, the two main points would be: successive classifications with removals of the classified classes and use the fewest classes possible as you suggested – Albert Sep 10 '15 at 7:05
  • @Albert C Those are certainly good ideas you mention. Often different bands contain some correlated information. A PCA (principle components analysis) can show how much "overlap" in information there is. In many cases much of the information, e.g. what spectral curve identifies an urban area, may be contained in a few bands, and the remainder are redundant. More on PCA here: help.arcgis.com/EN/arcgisdesktop/10.0/help/index.html#//…. – Dan Sep 10 '15 at 10:51
  • Regarding your second question in the first comment, I would mask out areas that you are not interested in before you do the analysis, as they may impact the classification. For example, you could mask or clip the image to an area around your key area of interest, e.g. removing open ocean, mountains, or other areas of the image that are not of interest. You could then use bands 764 (assuming you are using Landsat 8) to classify the image, perhaps selecting slightly more classes than you need. Then, create ROIs and run a classified analysis method for the number of classes you would like. – Dan Sep 10 '15 at 11:01
3

It is opinion not the answer. No room in comments box, sorry.

I’d say it is all about phenomenon of interest and expertise. There are 2 questions to answer:

  • number of categories you are trying to identify
  • list of relevant variables/rasters that can help.

In this case dealing with known number of vegetation classes it seemed reasonable to include info to account for air temperature (DEM) and moisture availability (tabular data on snow water equivalent vs elevation and aspect => raster). Proximity to known classes or smart translation of soil classes to numbers might help as well. This will produce defendable results. However adding more layers to the pile we’ll make a model unstable. Again, it is my opinion, it is how I feel about it.

But when I see a rainfall pattern as a factor in predicting urban development over tiny area, I am at loss with words. Again, it is my opinion, it is how I feel about it.

At the end of the day it is statistics at work. This is why, I think the appropriate list of not-correlated bands is A priority here. After this try, try, try.

  • Thanks for your reply!! Adding DEM, aspect, solar radiation and other sources is a wonderful idea to narrow down the classification in some cases. I like what you have done with the terrestrial landscape photographs. I couldn't access the paper but I find the idea itself very interesting. – Albert Sep 10 '15 at 6:40
  • Sorry, I see now the paper wasn't yours. Anyway, I found your idea very helpful, thanx. – Albert Sep 10 '15 at 7:08

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.