2

I am working on Land Use/Land Cover Classification of the Hindon river(India) watershed extracted using Aster DEM. Landsat 8 scenes had been used and mosaicked for the required AOI using the Watershed shapefile. I am facing a lot of troubles in the Unsupervised Classification, ISODATA method as I am getting numerous mixed pixels which are difficult to label. I had also tried histogram equalization before clustering but the recoding & labelling issues persist. Kindly suggest me some procedure to overcome the difficulties faced in labelling the classes.

  • 1
    could you mention what software/version you are using? – mr.adam Jul 1 '15 at 16:00
  • I am using ERDAS and can use Arc GIS as well – Ashish Agarwalla Jul 2 '15 at 4:03
1

If you use QGIS, there is a plugin named "semi automatic classification". It may not take too much time to utilize the plugin because you might be familiar with the RS analysis methods. I have used it for 1 week and have been pleased. The plugin is also capable of downloading landsat's photos. Here's the link of classification tutorial. https://youtu.be/nZffzX_sMnk

  • I am using ERDAS and can use Arc GIS as well – Ashish Agarwalla Jul 2 '15 at 4:04
1

I have also had a lot of success using the image classification tools in ArcGIS (http://resources.arcgis.com/en/help/main/10.2/index.html#//00nv00000008000000). The documentation is great and the results have been very accurate.

Unsupervised classification is tricky because defining the number of classes will always result in some degree of mixing. Even over fairly uniform land cover types (e.g. forest or water), there is always some degree of variability. If you are concerned that a specific area seems mixed (i.e. a forest area contains different land cover types), then it might be helpful to inspect the spectral signature of each class. Are they spectrally similar? Could they perhaps be merged into a single class or are they distinct? Do you need more or less classes? These are the types of questions that you need to ask yourself.

Once you start probing the differences/similarities between classes, you might begin to see that there are several distinct land cover classes in your data. At this stage you might want to consider trying a supervised classification. For this you can use ArcGIS, ERDAS, or the QGIS tool that the previous answer recommended.

You might also consider running a majority filter on your initial classification (http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//009z00000037000000.htm). This will remove some erroneous pixels by merging them with their closest neighbors.

Last piece of advice. Even though a few pixels may seem misclassified or "mixed" when compared to high res orthos, running an accuracy assessment (e.g. a confusion matrix) will allow you to state the accuracy of your classification. No classification is perfect, so it is important to state its accuracy.

Good luck!

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.