I am not following what the problem is here. The rule images from ENVI's Maximum likelihood classifier is in chi squared values, not probabilities, so you can have negative values. You can enter probability values in the rule classifier and ENVI will convert them to chi square values to provide the updated classification. You should not need to stretch the rule images to get a good, new classification from the rule classifier.

Thank you, MariM, this problem with probabilities was already highlighted here in forum. Can i also use ROC CURVES for probabilities?

ROC curves do use probabilities to visualize a classifier's performance to aid in selecting a proper threshold to use in the Rule Classifier. Here is a good description of how they can be used: ROC curves are used to visualize a classification’s performance and to help in selecting a suitable operating point. For example, choosing they optimum distance threshold for a minimum distance classifier. They are created by plotting the probability of detection versus the probability of false alarms for a series of classifications. The selected operating point then depends on the desired probability of detection (PD) and probability of false alarm (PFA). Consequently, ROC curves may show that in some cases a selected classification method cannot perform to the desired PD and PFA. The following steps illustrate the processes ENVI uses for calculating a ROC curve: 1. Compute a threshold and classify the rule image based on this threshold 2. Compare the classified image to the ground truth and count the number of pixels classified properly and the number of pixels classified improperly. 3. Compute one of the points for the ROC curve where: PD = points classified properly / # of points in class PFA = points classified improperly / # of point not in class 4. Repeat number 1-3 for a number of different threshold values. Each repeated point becomes a point on the ROC curve.

Much thanks! I have used this way: 1. I took my classified image ( Rule image) and used it to create ROC Curve 2. Then, according to require for ground truth in display, i choose my reference data 3. Then i marked all classes in display (in my case-10 classes) for ROC Curves 4. Then i gave min and max. value for x-achse (in my case 0-40) 5. Finally i got 2 windows (each window with 10 curves), in one of them i got false alarm- versus PD, and in other window- curve for threshold. 6. I have adapted for example 99,5% in window with truth/false alarm and choosen then in window for threshold- THRESHOLD-VALUE ACCORDING to this 99,5% VALUE in Y-achse. I hope, it is correct

Dear MariM I have two questions: 1. For optimum operating point, I selected point of minimum distance from (1,0) and which has minimum 0.9 PD. Is it correct? Are there other methods? Any reference paper for this? 2. For maximum likelihood, I always get a constant line on PD = 1 in ROC. How to solve this?

I cannot confirm that your min/max threshold values that you specify for the ROC analysis are appropriate for your data. I would calculate statistics on your rule images, and see what the range of values are, and use something related to the range of values in the rule images. Other than that piece, this sounds right to me. For the first class, it soundsl ike you have a probability of detection around 1 for any threshold. So correctly detecting is not sensitive to the specific threshold used. I sometimes see this, if the first class is something quite spectrally different from the other classes. You should see some variation in the probability of detection for the other classes, though, if things are set up correctly. A reference for the ROC method is found in the ENVI help for ROC, and is: A. P. Bradley, 1997, “The use of the area under the ROC Curve in the evaluation of machine learning algorithms,” Pattern Recognition, V. 30, No.7, pp 1145-1159. Peg Exelis VIS

Oh that really helped. Thanks a lot Peg :)

Dear Peg I did as you said to change percent values in rule classifier for maximum likelihood. I have obtained the waterfall curve in ROC through which I selected the optimum operating point (threshold) with high PD and low PFA. But the problem is that the threshold range is from 0 to 1 for MLC classifier and the optimum points I get from ROC are above 20. Will I have to convert DN to reflectance or what? Regards

Ah, yes. It is true that if you go back to the original tool you used for the Maximum Likelihood classification, you'll need to enter your threshold in terms of probability (between 0 and 1). And the threshold values you get from the ROC tool are in the units used in the rule images, which are not probability values, they are chi squared. So, you will instead want to use the Rule Classifier tool in ENVI. When you enter your new threshold values (from the ROC tool) there, it will accept them. - Peg

Dear Peg Thank you for your support. I did it with your help. You rocked (y)