Objectives The combination of spectral and spatial information can significantly improve the classification performance of hyperspectral images. However, the number of labeled training samples is limi‑ted in real scenes, which restricts the classification accuracy. To make full use of spectral spatial information and solve the problem of limited number of hyperspectral image labeled samples, we propose a hyperspectral image classification method based on nested sliding window sample augment and spectral space iteration.

Methods First, principal component transform is used to reduce spectral dimension of hyperspectral image. Then, a sample augment method based on nested sliding windows is proposed, in which representative pixels are selected from the optimal sub windows and added to the training set. Finally, iterative support vector machine and joint bilateral filter are used for classification, and a decision rule is proposed to terminate iteration to obtain the final classification diagram.

Results It is verified on three hyperspectral datasets: Indian Pines, Pavia University and Salinas. The experimental results show that when the training samples are limited and the samples of the three datasets are 2%, 0.2% and 0.2% respectively, the classification accuracy of the algorithm reaches 94.42%, 95.82% and 96.39% respectively. Compared with spectral classifier support vector machine, the overall classification accuracy of the proposed algorithm on the three datasets is improved by 28.1%, 20.1% and 13.1% respectively. Compared with S³-PCA based on small samples, the overall classification accuracy of the proposed algorithm on the three datasets is improved by about 3%, 5.5% and 4.5% respectively.

Conclusions The proposed algorithm can obtain satisfactory classification results in the case of limited training samples, which is superior to other advanced comparison algorithms.