Dear all, my research focused on textural analysis using GLCM to enhance classification accuracy of land use of tropical rainforest in Borneo. I create several texture image and then combine it with spectral image as input data in the process of image classification. Mathematically, how do I know if texture measure can improve accuracy of image classification?
Thank you so much.
1 Using same training and testing dataset
2 Evaluation of classification result only using spectral image
3 Evaluation of classification result using both texture and spectral image
4 If result of 3 is better than that of 2, then we can say the addition of texture can help improve the classification accuracy
Although you can proceed as Lijun pointed out, keep in mind that this not prove your point since such analysis is also dependent of your dataset and the intrinsic features (not the ones you get from GLCM) it has.
Thanks you all for your answers. I have read some papers concerned about it. Most of the researcher used accuracy (using error matrix method, Kappa index, etc) as indicator to prove that texture image can improve accuracy of image classification. Although we actually knew that there were parameters influenced accuracy like geometric correction, training samples, etc. What are the approaches can I use to prove that additional of texture image in image classification can improve classification accuracy? Thanks...
My two pennies. There are two ways to make a judgement on whether or not certain features impact on class separation. One is making experiment -- that would involve features you compare against, as well as certain pool of classifiers of your choice. This would not preclude you from being affected by your choices (your samples, cross-validation methods, classifiers or else). Another way is justifying separability based on some formal criterion (like linear discriminant measure, Pearson correlation coefficient, or plethora of other algorithm-specific criteria). One important consideration here is that your judgement is not specific to actual classification -- that may be considered a drawback or benefit, depending on needs and situation.
There is one catch to the seemingly obvious method proposed by Lijun: it only gives you an indication of the benefit of features if the number of training examples exceeds the number of features by orders of magnitude. If not, you may encounter the "curse of dimensionality", reflected in randomly correct class assignments. I'd suggest to look at the class distribution histograms of individual features, and of course follow Nikitas advise after a thorough read of the according literature.
Best wishes for your endeavours!
- Badges
- Science topic