Research On Classification Of Hyperspectral Images Based On Active Learning And CNN

Hyperspectral image classification is a hot research issue in the field of remote sensing image processing.However,it is difficult to obtain a large number of labeled samples in hyperspectral image,and the cost of manual labeling is large,which limits the improvement of classification performance.To solve these problems,classification methods for hyperspectral image based on active learning and convolutional neural network are designed in this thesis.Through active learning,valuable samples are selected to improve the performance of classification model.In the case of less training samples,it can achieve better classification results.The main contents of this thesis are as follows:First,because of the lack of labeled samples in hyperspectral image classification,the performance of classification is limited,a hyperspectral image classification method based on active learning is proposed in this thesis.Using the rich spectral information of hyperspectral image,the method is constructed from two aspects of spectral and spectralspatial.The training set is purposefully constructed by active learning,which provides valuable samples for convolutional neural network.And different sampling strategies are used in the experiment.The experimental results show that active learning method can get better classification results than random sampling method.Second,in hyperspectral image classification,the traditional active learning only uses the data labeled by experts as training samples,and a large number of unlabeled data are not fully utilized.A hyperspectral image classification method based on active deep learning is studied in this thesis.The spectral band of hyperspectral image is analyzed by principal component analysis,and the spectral-spatial information of the image is used.The convolutional neural network is trained by the labeled samples and the probability of the unlabeled samples is predicted.The sampling strategy is used to evaluate the samples.For the samples with high information,experts give them labels;and for the samples with high confidence,the computer automatically gives them prediction labels.The experimental results show that the proposed method in this chapter is better than the random sampling method in the same labeled sample size.Third,in view of the problem that the single strategy is not comprehensive enough to evaluate the value of sample,and it is easy to introduce false labels,a hyperspectral image classification method based on Multi-Strategy-Check active deep learning is proposed in this thesis.This method is improved in the core part of sample selection.In order to select unlabeled samples with high reliability,this method combines a variety of sampling strategies to select samples by judging the difference and certainty of samples.The experimental results show that this method can achieve high classification accuracy and greatly reduce the cost of manual labeling.