Keywords: fused image, test site, verification
Abstract In this paper, in order to improve the classification accuracy training data screening and truncated normal distribution maximum likelihood classification are adopted. We verify these techniques are effective measures for the improvement in the classification accuracy.
1. Introduction
Generally classified result is changed by how to get the training data in the land cover classification using unsupervised algorithm. As the training data are extracted by the human interpretation, pixels including errors are extracted. Therefore, it becomes the classification accuracy lower. By removing the extracted pixels including errors, the classification accuracy will be improved. The normal distribution is assumed for the distribution of each class in the maximum likelihood classification. But actual distribution is different from normal distribution. The misclassification decreases by removing the tail of each class in which the misclassification seems to increase, and the classification accuracy will be improved. In this paper, in order to improve the classification accuracy training data screening and truncated normal distribution maximum likelihood classification are adopted. We verify these techniques are effective measures for the improvement in the classification accuracy. The distribution profile of training data is examined. Probability density function of each class is recalculated after removing the tail of each distribution. In other words, the distribution with large standard deviation converts into small distribution profile. Instead of the normal distribution in the case of maximum likelihood method, truncated normal distribution is used in truncated normal distribution maximum likelihood classification. Classification accuracy will be calculated by comparing test site data and classified result. As a result of the experiment, it is confirmed that screening of training data and truncated normal distribution maximum likelihood classification are effective measures for the classification accuracy improvement.
2. The screening of the training data
With the screening of the training data, the distribution profile of normalized training data is examined, and probability density function is recalculated after excluding the pixels, which cause misclassification. There are two methods for screening.
(1) The pixel was removed when pixel value exist in the tail of training data distribution at least one band.
(2) The pixel is removed when pixel value exist in the tail of training data distribution at all bands
3. Maximum likelihood classification using the truncated normal distribution
3.1 Outline of the truncated normal distribution
Truncated normal distribution is made from normal distribution by truncating the tail of the distribution. The probability distribution function of truncated normal distribution is defined in equation (3.2) using A(a) of following equation (3.1).
In the maximum likelihood classification normal distribution is used for the distribution of the population of each class. Classification result differs by the distribution profile. Generally, a region of the class from mean value of each class is decided by standard deviation. The region of each class is decided by the coefficient of this standard deviation s. The classification accuracy will be improved by removing the part of the tail of normal distribution in each class. The part of the tail is mixel region of the distribution profile, and the misclassification will be decreased. In other words, by accurately classifying the part of the tail, the improvement of the classification accuracy can be expected. The number of undiscriminant pixels increase, when the coefficient of s is decreased. The number of undiscriminant pixels decrease, when the coefficient of s is increased.
4. Experiments
Object image used in this study is fused image of panchromatic image of SPOT and 2,3,4 band of Landsat/TM. The classification was carried out in 50 classes, and each category was finally integrated to 7 classes, housing area, grassland, paddy field, coniferous forest, bare ground, water body and shadow.
