Abstract
The objective of this study is to propose a method to purify the training sample for the Maximum Likelihood Classification (MLC). Furthermore, we discuss whether purifying the training samples with the proposed method is significant or not for improvement of classification accuracy. The procedure of the proposed method consists of five steps as follows:
STEP-1) Preparing the preliminary designated training data through the detailed ground truth.
STEP-2) Classifying the preliminary designated training data by using MLC.
STEP-3) Clustering the preliminary designated training data by using ISO-DATA algorithm.
STEP-4) By using the separability as the index of the distance between classes, one to one correspondence is found between each classes acquired through the processing of step-2 and step-3 respectively. After that, the pixels which belong to the same class are selected to purify the training data.
STEP-5) Classifying the pixels based on the class statistics of the purified training data by using MLC.
The effectiveness of the proposed method were verified for the HRV data. The summaries of the results are as follows:
1) The proposed method can purify the training data automatically and effectively based on the statistical proceeding.
2) In comparison of the statistic values of the purified training data with those of the preliminary designated training data, the separability in terms of divergence was improved among all pairs of classes. Furthermore, the goodness of fit to the multivariate normal distribution was also improved.
3) The accuracy of PCC (Probability of Correct Classification) was increased to 92.6% from 84.9%.
The proposed method for purifying the training data in this study is very practical and useful for classifying high resolution satellite data in case of using MLC.
