Abstract:

Investigating a new clustering methods is always paid more attention by the researchers in statistics and machine learning fields. Noticing that probabilistic distance clustering cannot solve nonlinear clustering problems, the authors proposed a new nonlinear clustering method called kernel probabilistic distance clustering by extending probabilistic distance clustering algorithm with kernel function theories,and then laid out an executing strategy and feasible algorithm of the new model for the unsupervised classification of remotely sensed image data,and finally developed a visual C++ program of the new model for remotely sensed data classification on the basis of GDAL function library for the input and output of remotely sensed data. A TM image with seven bands and 491× 440 pixels in size provided by ERDAS software platform was applied to the experimental application of the new model for unsupervised classification. The authors compared the clustering performances of the new model and its original version in the unsupervised classification of the TM images. The experimental results show that the clustering performance of new model is better than that of the original linear version in unsupervised classification.

Key words: probabilistic distance clustering, kernel function, kernel probabilistic distance clustering, remotely sensed images, unsupervised classification