Statistical models for spatial multivariate date and discriminate analysis

点击翻译按钮获取中文摘要

The main aim of this study was to derive an efficient classification procedure for land-cover categories based on geospatial date. Assuming normal distributions for feature vectors and a structural Markov random field (MRF) for category distribution, we obtained a classification method, which shows an excellent performance for real date. Then, we proceeded to examine general structural MRF and an estimation method of unknown parameters. These results were presented as invited talks at two international conferences (submitted for publication).It was shown that MRF models the spatial dependency of the categories well. Next, we take machine-learning approach for feature space. Probabilistic support vector machine (SVM ) is treated and the posterior probability is defined by the loss function. This approach gives a similar efficiency due to the statistical approach. These is, however, a room for improvement of the posterior. This point is now under investigation.Another attention was paid for microarray data on human gene. The aim is to classify cell types by gene expression patterns through machine learning. First, dozens of genes are selected among several thousands genes by AdaBoost. Then, SVM chooses the best combination of the selected genes. Our approach is highly efficient than the ordinary (in preparation).