Study Of Dimensionality Reduction And Spatial-spectral Method For Classification Of Hyperspectral Remote Sensing Image

HyperSpectral Image(HSI)classification is an important problem in the study field of remote sensing.Hyperspectral data provide abundant spectral and spatial information,which has a beneficial effect on fine identification of terrain.The accuracy of classification will be affected by the quality of classification methods.The spatial-spectral classification method which use the spectral features and spatial information together,has been proved to improve the classification accuracy significantly.Dimensionality reduction of hyperspectral data can save time and avoid dimension disaster,and it has become a major step of classification.Depth learning can extract essential feature from data and can become an important research topic for classification of HSI now.This paper studied information theory,machine learning,depth learning and other fields to improve the accuracy of classification.The main work is summarized as follows:Two novel feature selection methods are proposed according to two evaluation criterions: information quantity and interclass reparability.One innovative method is an Improved Optimum Index Factor(IOIF)method.Values distribution of the correlation coefficient matrix of HSI have obvious “block” characteristic.These blocks can be used to divides all bands into several subspaces.Then IOIF is calculated for each bands group which comes from different subspaces.Finally,band selection is carried out according to IOIF values of these groups.This method improves the computational efficiency and the classification accuracy,and reduces the information redundancy.The other innovative method is a novel binary particle swarm optimization with mutation mechanism(MNBPSO)algorithm.The MNBPSO method is applied to feature selection and classifier parameter synchronization optimization.The experimental results show that the proposed algorithm works effectively,and improves the HSI classification accuracy.Based on the ability of deep learning algorithms to extract the essential features of data,a stacked denoising autoencoders(SDAE)network model is optimized for HSI feature extraction and classification.The experimental results show that SDAE works better than some conventional feature extraction methods.Two spectral-spatial HSI classification methods are proposed based on Mathematical Morphology.One is the spatial post-processing method based on binary mathematical morphology.This method utilizes the noise reduction ability of binary morphology to optimize the classification results based on spectral features,which greatly reduces the salt and pepper phenomenon in the classification results and significantly improves the classification accuracy.The other is the spatial synchronization processing method based on gray mathematical morphology.In this method,each band of hyperspectral data is transformed into gray image,and then denoised separately.The achieved spatial-spectral features are used as the input of the above optimized SDAE classifier.In this paper,we have put forward several innovative points from the two directions of dimensionality reduction and spatial-spectral classification.Some synthetic methods are proposed based on these innovative points.Our proposed algorithms are compared with several similar state-of-the-art methods.Experiments show that our proposed approaches could achieve better classification accuracy than these state-of-the-art methods.