Research On Semi-supervised Classification Method Of Hyperspectral Images Based On Convolutional Neural Network

Hyperspectral image(Abbreviated as HSI)can not only provide detailed surface material information,but also contain a large amount of spectral information,spatial information and radiant energy.Therefore,researchers have gradually begun to attach importance to the development of hyperspectral imagery.Hyperspectral data has hundreds of thousands of dimensions and contains rich spatial and spectral information,but it is for this reason that it is easy to cause dimensional disasters and data redundancy.Therefore,it is very important to find a suitable method for process hyperspectral remote sensing data.As deep learning has obtained good effects in the fields of computer vision and image processing,current methods of deep learning have also been widely used in the field of hyperspectral image classification.Among them,Convolutional Neural Network(Abbreviated as CNN)is widely used.When CNN classifies hyperspectral image in the spectral domain,it can learn highly representative image features from the training data.In addition,because there are few labeled samples of hyperspectral data,the number of available training samples is very limited,and a large amount of unlabeled data cannot be used.In fact,these unlabeled samples also have great value in classification tasks,so a semi-supervised algorithm is proposed.The semi-supervised algorithm on the one hand makes use of the limited supervised information transmitted by labeled samples,on the other hand it also takes into account the rich information contained in a large number of unlabeled samples.Therefore,it is of great significance to improve the classification performance.This paper combines the spectral and spatial information of hyperspectral image to researches the classification method.The main research contents are as follows:(1)Aiming at the phenomenon of same-spectrum foreign matter and foreign matter homospectrum in hyperspectral image,a semi-supervised hyperspectral image classification framework based on metric learning is proposed.The proposed discriminant objective function based on metric learning can improve the classification results by reducing the distance between classes and increasing the distance between classes.Finally,Markov random fields are combined to further utilize spatial information.Experimental results showthat on the Indian Pines and Pavia University datasets,this method has better classification performance.(2)Aiming at the problem that it is difficult to train the imbalanced data samples in hyperspectral image,a semi-supervised hyperspectral image classification framework based on focal loss is proposed.The proposed multi-class objective function based on focal loss can reduce the weight of easy-to-divide samples and increase the weight of difficult-to-divide error-prone samples.Finally,considering the impact of spatial information of hyperspectral image on classification,Markov random fields are used as image post-processing.Experimental results show that the method is superior to other deep learning methods on the Indian Pines and Salinas datasets.In summary,combined with the scarcity of hyperspectral data training data,the method of hyperspectral image classification proposed in the paper can fully combine spectral information and spatial information to achieve a more accurate hyperspectral classification.