Hyperspectral Image Classification Based On Deep Learning

Hyperspectral images have high spectral and spatial resolution.Recently the analysis and processing of hyperspectral images has become one of the hotspots in the research of remote sensing image.Hyperspectral image classification task plays an important role in the fields of geological exploration,crop detection,national defense and military.Therefore,it deserves more in-depth research.Deep learning has excellent ability to extract the essential features of data.To improve the classification accuracy of hyperspectral image,two deep learning-based methods for hyperspectral image classification were proposed in this thesis.The main research contents are as follows:(1)A double-branch coupling network used for dimensionality reduction of spectral vectors is proposed in this thesis.The network is consisted of one-dimension convolutional coupling network and word embedding coupling network,which can realize the mapping from the category space to the dimensionality reduction representation space and can introduce clustering characteristic for the dimensionality reduction output.Experiments show that the introduction of clustering characteristic increases the discriminativeness between classes and reduces intra-class differences,which can decrease the misclassification brought by "same object with different spectrum,different objects with same spectrum" and improve the hyperspectral image classification results.(2)A loss function and an iteratively alternate training method based on greedy strategy are proposed for the double-branch coupling network.The loss function includes classification loss and clustering loss used to introduce clustering characteristics for the dimensionality reduction output.According to this method,the classification loss and the clustering loss are alternately minimized.By iteratively finding the local optimal solution,the parameters of the network can gradually reach the global optimal solution.Experiments show that a dimensionality-reduced representation that is conducive to classification can be obtained based on the proposed training method.(3)A hyperspectral image classification method based on the generative category probability distribution is proposed,in which a gaussian distribution is used to represent the category information of the training samples to obtain the category probability distribution vector samples.The vector samples can be used as real samples for the training of the generative adversarial network.The generator can generate a category probability distribution based on the learned real sample distribution to achieve the classification of hyperspectral images and the discriminator only needs to complete the true and false discrimination during the network training.At the same time,A semi-supervised network model is proposed to learn from the class probability distribution vector samples and unlabeled samples.Unlike the most GAN-based classification methods which input the category information of training samples as conditional information into the network,in this semi-supervised network model,features extracted in an unsupervised manner are used as conditional information.Thus the unlabeled samples can be used for the feature extraction.The experimantal results on three HSI data sets verify that this method solves the problem of the discriminator’s distinguishing the data source and classifying the samples simultaneously during the network training when the generative adversarial network is applied to multi-classification tasks and can effectively improve the classification result.