How to Create Classifiers

Creating classifiers always proceeds along the same lines:

• Define the spectral descriptors. Any of the statistical modelling techniques used in Epina ImageLab does not use intensities/absorbances/abundances directly but always uses spectral descriptors as the variables describing a data set. Please note that spectral descriptors include plain intensities as well.
• Define the training set. The training set should be comprised of 10 to 100 sample spectra for each class. Depending on the goal of the classification (simple yes/no distinction, or multi-class discrimination) the training set should be selected in a manner, that all typical cases for the classes to be discriminated are included. Epina ImageLab supports the creation of a training set by providing a Test and Training Set Editor.
• Define a test set. The test set should be selected under similar constraints as the training set. If possible the test set should be of similar size as the training set. It must not contain any image points which are already part of the training set (use the Test and Training Set Editor to create a test set).
• Create the classifier using a particular method. Currently, Epina ImageLab supports the following classification methods:
  
  • Discriminant PLS
  • Random Decision Forest
• Test the classifier. Thorough testing of the classifier is of crucial importance. In principle, the results of both cross validation and the application of the classifier to a test should be considered. Further, the classifier stability against noisy data should be checked.
• Use the classifier and/or release it for general use. Depending on the type of model you may use one of the built-in applicators:
  • PLS/DA Model
  • Random Forest Model