平成17（2005）年度　一般研究２実施報告書

課題番号	17−共研−2033			専門分類		6
研究課題名	空間における関数データの統計的問題の研究
フリガナ 代表者氏名	ニシイ　リュウエイ 西井　龍映			ローマ字		Nishii Ryuei
所属機関	九州大学
所属部局	大学院数理学研究院
職　　名	教授
所在地
TEL				FAX
E-mail
URL
配分経費	研究費	0千円	旅　費		0千円		研究参加者数	5　人

研究目的と成果（経過）の概要

In recent years, hyperspectral imaging sensors are paid attention because
the plentiful number of bands give us much spectral information. The sensors observe the ground surface and acquire digital images of hundreds of spectral bands. Hyperspectral images can, therefore, be used for deriving more accurate discrimination than conventional multispectral images.
Over the past few decades a considerable number of studies have been made on statistical classification for multispectral imaging sensor. We will pursue efficient classifiers for hyperspectral images. The methods are not useful, however, if they need much computational time in building classifiers or classifying test data. A simple dimensional reduction method, such as feature selection, gets rid of this problem, but loses much spectral information. Hyperspectral data should be, therefore, discriminated in high accuracies without losing almost spectral information under the acceptable computational time.
Support vector machine(SVM) and Neural network(NN) are mentioned as current discrimination method for hyperspectral data. These are powerful predictive methods, but very complex models. In addition, a user faces to parameter decision problem. Here, AdaBoost is employed for hyperspectral image classification. AdaBoost combines weak classifiers adaptively and discriminte targets by a weighted committee of weak classifiers. By considering simple weak classifiers, AdaBoost is superior in interpretability for extraction of useful features for a discrimination than SVM or NN. We propose multistump AdaBoost and it is applied to AVIRIS data. We show the method to escape numerical problem which arises from adaptation for hyperspectral images. It gives high accurate results than SVM or NN does. It is also applied to multispectral data. This study is under preparation for publication.

Our publications in 2005-2006 are listed in F1-F4 and P1-P3. They are related to supervised image classification based on statistical machine learning. Spatial AdaBoost was proposed by F1. Then, it was extended by P1 and P2 in various settings for applications. F2 introduced new Markov random fields (MRF) useful for image classification. F3 is a review paper of our recent works. F4 proposed contextual clustering methods based on MRF.

当該研究に関する情報源（論文発表、学会発表、プレプリント、ホームページ等）

F1. Nishii, R. and Eguchi, S. (2005).
Supervised image classification by contextual AdaBoost based on posteriors in neighborhoods.
IEEE Transactions on Geoscience and Remote Sensing, 43(11), 2547-54.

F2. Nishii, R. and Eguchi, S. (2006).
Image classification based on Markov random field models with Jeffreys divergence.
to appear in Journal of Multivariate Analysis.

F3. Nishii, R. and Eguchi, S. (2006).
Supervised image classification of multispectral images based on statistical machine learning.
to appear in Signal and Image Processing for Remote Sensing, Edited by C. H. Chen.

F4. 川口修治, 山崎謙介, 西井龍映 (2006).
ミクセルを考慮したマルコフ確率場に基づくリモートセンシング画像の教師なし土地被覆分類.
リモートセンシング学会誌に掲載決定（26巻2号の予定）

P1. Nishii, R. and Eguchi, S. (2005).
Spatio-temporal contextual image classification based on spatial AdaBoost.
Proc. of IEEE International Geoscience and Remote Sensing Symposium, CD-ROM.

P2. Nishii, R. and Eguchi, S. (2005).
Robust supervised image classifiers by spatial AdaBoost based on robust loss functions.
Proc. of SPIE, Vol. 5982, 59820D.

P3. Tanaka, S. and Nishii, R. (2005).
Verification of deforestation in East Asia by spatial logit models due to population and
relief energy. Proc. of SPIE, Vol. 5976, 59760W.

研究会を開催した場合は、テーマ・日時・場所・参加者数を記入してください。

研究参加者一覧
氏名	所属機関
江口　真透	統計数理研究所
小西　貞則	九州大学
竹之内　高志	奈良先端科学技術大学院大学
松浦　正明	癌研究会