I am trying to use Landsat data to map urban extents over time. I need to perform a classification into built up / non built up areas. I am new to image classification. I have access to the following software:

  • ArcGis
  • ENVI
  • Ecognition
  • GRASS / any other open source software you can recommend.

Which methods and applications are most suited to this analysis? I would appreciate if you could help me understand which methods or applications yield the highest classification accuracy.

  • 1
    I would appreciate if you could help me understand which tool does a more accurate job. I heard that ArcGis is not great.... – Sheila May 1 '13 at 23:55
  • 1
    This question looks like going into an extended discussion/debate on pros/cons of software choice so I have voted for it to be closed. – PolyGeo May 2 '13 at 4:22
  • @PolyGeo The OP is clearly looking for an accuracy assessment or resources to guide the analysis. There is no reason to close this post. – Aaron May 2 '13 at 4:43
  • @Aaron Sounds like you inferred "best procedure" where I read "best software" literally. Unfortunately, I cannot unvote for closed so will have to leave it there. – PolyGeo May 2 '13 at 5:26
  • Thanks everyone for the comments. I am sorry if starting this discussion was not appropriate, I had no idea! – Sheila May 2 '13 at 14:48

Keep in mind - no one procedure is necessarily going to provide the "best result." Image interpretation is critical, both before and after classification. You will likely find urban areas misclassified as something else and non-urban areas classified as being urban.

You have two basic approaches:

1) Supervised classification: this involves selecting pixels from the image that you know represent urban areas. The selected algorithm will find pixels with similar spectral properties and assign them to the "urban" class. Note that urban areas are highly heterogeneous, meaning that any one "sample" may not be enough (mix of trees, roads, buildings, grass, etc.). You may want to select different types of urban areas and then aggregate them into one "urban area" class (i.e., begin by identifying urban residential, urban commercial, etc.).

2) Unsupervised classification: In this scenario, you choose a number of pixel "clusters" you want the algorithm to identify based on statistical similarity. You then examine each cluster and identify the type of land cover it represents. It is typically recommended that you choose a number of clusters which far exceeds the number of classes you eventually wish to end up with (you aggregate results). This has the advantage of helping you to identify the number of different urban settings that exist within your image (you may see one cluster of pixels that primarily makes up heavy industrial areas and another cluster that makes up residential areas). Note that these clusters will "overlap" in terms of class (i.e., several clusters may represent a single land cover type). If you find a cluster that includes pixels of more than one class (forest and urban residential), you may wish to create more clusters from that cluster - a technique known as "cluster busting."

Other notes:

| improve this answer | |

Let's narrow down the methods of classification to two major groups: object-oriented classification and pixel-based classification. The attached tables are from a publication titled Comparison of Pixel-Based and Object-Oriented Classification Approaches using Landsat-7 ETM sPECTRAL Bands. The highlighted row in Table 3 shows that object-oriented classification (e.g. image segmentation with eCognition) has the second highest producer's and user's accuracy in addition to the second highest kappa statistic for the settlement area information class. However, the object-oriented approach has the highest overall accuracy and overall kappa.

From my own experience, and as mentioned by others in this thread, eCognition has a very steep learning curve. Results are often difficult to repeat across multiple Landsat tiles. Automation is also a challenge unless you are familiar with eCognition Server. However, this is the gold standard for image segmentation software. In theory, you will likely obtain the best result from this method.

Based on this publication, I would also like to encourage you to consider supervised classification with a minimum distance decision rule. I have found image processing with ArcMap to be a bit of a black box. For example, you cannot specify which decision rules to use in ArcGIS supervised (maximum likelihood) classification, whereas you can specify these rules in Erdas Imagine.

My opinion is to start simple with supervised maximum liklihood classification and scale-up to eCognition if the need arises.

enter image description here

| improve this answer | |

While ArcGIS can certainly perform the work you want, if you have access to other software, I would use those. I have not used GRASS before for image classification, but I know it is more than capable. From my knowledge, Ecognition is more for object based classification. I have used ENVI for image classification a few years ago and it does the job really well. As far as which one is the most accurate, it's hard to say. I know that ENVI is a specialised raster processing software and has been around for years. GRASS, too, has been around for a long time and will probably do just as good a job.

It probably just depends on what you feel most comfortable with. ENVI is fairly straight forward and you could probably pick it up fairly easily. If you've never used GRASS before there is a bit of a learning curve. Opticks, too would probably work well, but again, there might be a learning curve.

There is a lot of science and art involved in image classification. If you haven't done much of it before, I would read up on techniques and theories first. You can spend a year at university learning about it.

| improve this answer | |


Might be another good option, looks quite functional.

| improve this answer | |

If you use ArcGIS you'll want the spatial analyst extension. The tutorials for image classification are pretty easy to follow, and the actual process is pretty easily followed with real world data (Landsat). Another source that may save you tons of work is from USDA. Google USDA NASS. The National Agriculture Statistics Service has a downloadable raster that is already classified, and web services. http://www.nass.usda.gov/

| improve this answer | |
  • Thanks! But I am looking into Asia, not the US, I doubt that NASS has a global raster? – Sheila May 1 '13 at 23:55

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.