Research And Application Of Hyperspectral Multi-classification Method

With the development of imaging spectrometers,the imaging precision of hyperspectral images is increasing dramatically.And hyperspectral imaging plays an important role in the field of economic construction and national defense.In contrast,the hyperspectral data dimension rises sharply and the nonlinearity increases,which poses a challenge to the traditional hyperspectral classification and the change detection method.This paper pays attention to existing problems in the traditional hyperspectral classification and detection methods and improves the accuracy and efficiency.This paper first improves the traditional OAO strategy in SVM classifier.By introducing the Chernoff distance and Jeffries-Matusita distance in the upper bound of the maximum error as the separability measure between classes,two different strategies are constructed respectively.At the same time,the kernel function of SVM classifier is customized,and the classifier is more adaptable to the data set by weighting.Experiments show that both methods can improve the accuracy and efficiency of classification algorithm effectively.Then Bag of Visual Word(BoVW)model for hyperspectral classification task is proposed.Hyperspectral data is complex,so BoVW model chooses CNN as the feature extraction part and overcomes the shortcomings of traditional classifiers effectively.The paper improves the SLIC algorithm at the same time,so that it can be used in hyperspectral image data.The BoVW model utilizes the spatial information to make up for the lack of spectral information in traditional method.The effectiveness of BoVW model is verified by experiments.Finally,the CNN-LSTM model is proposed for the hyperspectral change detection task.The change detection task can be regarded as a special classification task.The model uses the temporal information of the LSTM unit to improve the shortcomings of the traditional classification method.At the same time,features extracted by the proposed model ensure the accuracy of transfer learning when the samples are insufficient.The validity of the model is verified by normal change detection experiment and transfer learning experiments.It is proved that the model outperforms CNN model in the transfer experiment,and the importance of temporal information in hyperspectral change detection task is proved.