Research On Classification Algorithm Of Hyperspectral Imbalanced Data

Hyperspectral images have hundreds of spectral bands and are widely used in various fields,such as global environmental monitoring,land cover detection,resource management,natural disaster monitoring,interstellar detection,medical diagnosis and other applications.Through the classification algorithm,each pixel of hyperspectral data can be given a class label,which can extract useful information.It can be used in the extraction of thematic information,the production of thematic map,the detection of dynamic changes of ground objects,the establishment of spatial database and so on.Hyperspectral datasets usually have a serious problem of data imbalance,which will affect the classification accuracy of the single class,especially the class with a small number of samples.In practical application,minority classes are often the foreground targets that people pay attention to.However,the existing methods only focus on improving the overall classification accuracy of hyperspectral,and ignore the classification accuracy of minority classes.In order to solve the problem of data imbalance in hyperspectral classification,this paper has studied the hyperspectral classification algorithm.Based on the oversampling algorithm SMOTE and convolutional neural network,a new classification algorithm SMOTE-CNN is proposed.The oversampling algorithm SMOTE is used to generate artificial samples,next add them to the minority classes to make the training set reach the state of class balanced distribution,then use the training set of balanced distribution to train the convolutional neural network that has been built.Experimental results show that SMOTE-CNN can not only get high overall classification accuracy,but also get better classification accuracy of each class,especially in minority classes.Based on the cost-sensitive loss function and convolutional neural network,this paper proposes a cost-sensitive 1D-CNN for hyperspectral classification.In the training process of cost-sensitive convolutional neural network,a weight which isinversely proportional to the number of effective samples of the class is used to re-weight the loss,so that the network will tend to the minority classes in the training process.Therefore,cost-sensitive convolutional neural network can also learn accurate and robust features in the minority classes,then improve the classification accuracy of the minority classes.Experimental results show that,compared with the common convolutional neural network,the cost-sensitive convolutional neural network has greatly improved the classification accuracy in minority classes,and the overall classification performance has also improved.