Abstract:

In order to solve the problems of low classification accuracy and small number of samples in the process of hyperspectral image classification, an image classification method based on grid diversity and active learning is proposed. In this method, the principal component space is divided into several grids using the grid method. A sample is randomly selected from each grid containing samples, and the original spectral data is included in the training set. Then, using the active learning method, the K-nearest neighbor method is used to select some samples with the largest uncertainty among the remaining samples and incorporate them into the training set, thereby expanding the training set, making the data set representative, and improving the classification accuracy. In addition, in the process of data processing, principal component analysis and linear discriminant analysis are combined to reduce the dimension of spectral data, which further improves the operation speed. The experimental results show that in the Indian Pines hyperspectral data set,and in the case of a small number of training set samples, the proposed method improves the overall classification accuracy by 12.24% and 19.76%, respectively when compared with random diversity and non-active learning.

Key words: hyperspectral image, classification, grid diversity, active learning, K-nearest neighbor method, principal component analysis, linear discriminant analysis, small number of samples