High dimensional multi-spectral image classification by feature scaling for kernel fisher discriminant analysis
More and more difficulties are encountered when using Support Vector Machine (SVM) and kernel fisher discriminant analysis (KFDA) to classify high dimension multi-spectral image. In order to overcome these limitations, feature scaling for kernel fisher discriminant analysis (ES-KFDA) is used to classify the high dimension image. The high complexity computation of SVM is mitigated greatly and the performance of KFD in the presence of many irrelevant features is improved. Incorporated with wavelet transform in space domain and analysis of characteristics between spectra, we focus here on tuning the scaling factors of the feature scaling kernel by feature scaling for kernel fisher discriminant analysis, in a feature-scaling kernel, each feature has its own scaling factor. If some feature is insignicant or irrelevant for classication, the associated scaling factor will be set smaller; otherwise, it will be set larger. So the performance of ES-KFDA in the presence of many irrelevant features is mitigated greatly. Experiment results of AVIRIS 92AV3C show that the generalization ability of ES-KFDA is strong, and its classification accuracy is better than traditional algorithms, the training time of it is significantly shorter than SVM.