I'm new to this field of study.

my teacher ask me to do a hybrid MLC and ISODATA classification for my bachelor thesis. so I got one journal that combine both method, but I think it too complex for bachelor degree because it's also using a decision tree, which I don't understand at all.

so I was thinking, what if I made one for myself, a simple one but useful, accurate, and easy to apply. here how it goes:

I classify image with ISODATA, and then get the result, if accuracy assessment value is low, we redo the classification, if high, we use that result as training area for MLC. Then after we classify the image with MLC, we compare the result also from MLC but with training area from ground truth.

The conclusion will be, if both result are similiar, that mean we don't have to do ground truth anymore to get training area for MLC, we can just use ISODATA to get the training area and then use the result as training area for MLC.

I'm planning on doing all of this with ERDAS IMAGINE 2014

Can you give me your opinion for this method?

closed as primarily opinion-based by PolyGeo Jan 20 '18 at 20:57

Many good questions generate some degree of opinion based on expert experience, but answers to this question will tend to be almost entirely based on opinions, rather than facts, references, or specific expertise. If this question can be reworded to fit the rules in the help center, please edit the question.


Using an unsupervised classifier such as ISODATA to generate training data for a supervised classifier such as maximum likelihood will most likely lead to poor classification accuracy.

A typical pixel-based classification approach involves the following steps:

  1. Generate randomly stratified training data preferably with field samples obtained from high accuracy GPS device. Alternatively, generate training data through expertly selecting on-screen which correspond to the land cover classes.
  2. Apply the supervised classification algorithm
  3. Assess the results using an accuracy assessment (e.g. cross tabulation, kappa, measures of agreement). Often this is accomplished using a hold out sample of the training data (e.g. 30% of training data).

Getting back to the hybrid approach, it is often necessary to group the results of an unsupervised classification into more meaningful results. For example, your ISODATA classifier may determine three classes exist within pixels representing a lake. You would then need to group those three classes into one "water" class--a hybrid approach. That operation is easily accomplished using map algebra.

  • i see, so my method is one hella big flawed method. thank you for your advice. I'll look into it futhermore and discuss it with my teacher – Pija Giandra Jan 20 '18 at 3:53

Not the answer you're looking for? Browse other questions tagged or ask your own question.