Hyperspectral Image Classification Based On Active Learning

Supervised learning methods is generally adopted in hyperspectral image classification,but supervised learning methods require many training samples,while in hyperspectral images,there are few labeled samples and it is difficult to obtain many of labeled samples,resulting in low classification accuracy.Based on the active learning method,a hyperspectral image classification method to obtain high-precision classification results in the case of a small number of labeled samples is designed this paper.The main results of this paper are as follows:(1)Aiming at the problem of insufficient labeled samples in hyperspectral images,a method based on active deep learning for hyperspectral image classification has been proposed this paper.According to the convolutional neural network output mechanism,an active learning sampling strategy with maximum entropy and output difference is designed by this method.The sampling strategy is used to select samples that are valuable to the classification model,and the classification performance is improved.The results of hyperspectral image classification experiments show that compared with the supervised learning of random sampling,the active learning method can obtain higher classification accuracy in the case of a small number of labeled samples.(2)Aiming at the problem of uncertainty in sample prediction in hyperspectral image classification,a hyperspectral image classification method based on Bayesian active deep learning has been proposed this paper.Bayesian approximation inference is used to solve the posterior probability output of the convolutional neural network.According to the uncertainty of sample classification in posterior probability output,the maximum entropy and mutual information active learning sampling strategy is designed.The samples with uncertain classification results are selected by sampling strategy,then labeled and added to model training to improve classification performance.The experimental results show that compared with the supervised learning of random sampling,the proposed method can obtain higher classification accuracy under a small number of labeled samples.