Hyperspectral Image Classification of Convolutional Neural Network Combined with Valuable Samples

Aiming at the problem that the manual labeling of samples in the hyperspectral image classification is expensive and laborious, a large number of unlabeled samples are not effectively utilized and the classification results are not ideal. A method which can provide valuable samples and employ convolutional neural network to extract spectral spatial features for classification is proposed. Active learning method is used to construct a valuable training sample set by iteratively selecting the most uncertain samples through support vector machine which performs well in small sample classification, and labeling them. Then the 3D convolutional neural network is used to extract the spectral spatial features of hyperspectral image. The experimental results of the hyperspectral classification on Indian Pines and PaviaU datasets show that the proposed method (3D VS-CNN) is better than traditional classification methods.