On The Selecting Method of The Training Class for The Multispectral Classification.
スポンサーリンク
概要
- 論文の詳細を見る
The Representativeness and performance of the training data are the important factors to discuss the accuracy of the supervised classification for the satellite multispectral data.<BR>In this study, a selecting method for improving the representativeness of the training data, which is called IMR Metod (selecting method for IMproving the Representativeness of training data using nPDF algorithm), was proposed. This method is based on the nPDF (n-Probability Density Functions, H. Cetin, 1991) algorithm which is very useful for multi-dimensional data transformation and reduction. The procedures of IMR Method consist of five steps as follows :<BR>STEP-1) Preparing the preliminary designated training data through the detailed ground truth.<BR>STEP-2) Classifying the data with Maximum Likelihood Classification (MLC).<BR>STEP-3) Displaying the nPDF Plot of the data distribution in each classes.<BR>STEP-4) Making the GST (Guide image for Selecting Training class) image for selecting new training classes, which are represented with different colors in accordance with nonoverlapping and overlapping areas between classes in the nPDF feature space.<BR>STEP-5) Appending new training classes and reclassifying with MLC.<BR>The effectiveness of IMR method are verified for the HRV data. The summaries of distinctive features applying IMR Method are as follows :<BR>1) Through GST images, everyone can easily find out classes with large variance in multi-dimensional feature space and append new classes to the preliminary designated training classes.<BR>2) In case of applying the IMR Method, the accuracy of PCC (Probability of Correct Classification) was increased to 84.9% from 80.2%. This means that the representativeness of training data was improved.<BR>3) Furthermore, the iterative procedure for appending the training classes using IMR Method give asignificantly higher classification accuracy.<BR>4) For improving the performance of the training data, the IMR Method (resampling method for IMproving the performance of training data using maximum likelihood method and iso-data algorithm) was already proposed by ourselves in 1993. In case of combining IMP Method with IMR Method, it was confirmed the accuracy of PCC was increased to 92.6% from 80.2%, that is to say, the best classfication accuracy was achieved.<BR>In conclusions, this combined classifers proved to be a indispensable and practical technique for the improvement of the multispectral classification accuracy.
- 社団法人 日本リモートセンシング学会の論文