空間的一様性の検定を用いる多重分光画像の教師なし分類法

抄録

A new clustering method to be used for multi-spectral images is proposed. This method consists of division and merging. The division uses a segmentation technique, in which a multi-spectral image is divided into spatially uniform areas. The division rule is based on the (spatial) Mahalanobis distance between coefficient vectors of a local regression model fitted to the neighboring areas of the image. As this distance is a chi-square statistic, a statistical test is employed to evaluate the significance of the distance, and the threshold for the test is theoretically derived.
The merging uses a clustering technique, in which the divided ares are merged into clusters. The merging rule is based on the (spectral) Mahalanobis distance between mean vectors of the multi-spectral data in divided areas. As this spectral distance also follows the chi-square statistics, the same statistical test is used. A threshold is determined for the test.
In this paper, we describe the principle and the procedures of the test. The superiority of this method to a pixel-by-pixel method with respect to the accuracy and the processing speed is confirmed quantitatively by numerical simulations.