Hyperspectral Image Feature Extraction Algorithm Based On Deep Matrix Decomposition And Its Applications

In the classification problem of hyperspectral images,obtaining labeled samples needs high labor cost.Therefore,the number of labeled samples in a hyperspectral image is limited.The high feature dimension and the information redundancy of hyperspectral images bring great challenges to classification problem.Therefore,effective feature extraction is an essential pre-process before classification.Through an efficient feature extraction,the redundant and noisy features are removed effectively and the favorable features are retained.Traditional feature extraction algorithms mostly extract shallow and low-level features,which has limited the ability of feature representation.Deep learning can express information hierarchically through deep structure,i.e.,we can extract richer features from images.Based on deep matrix decompositions,this dissertation proposes two feature extraction algorithms for hyperspectral images to improve the classification performance.The specific research contents of this dissertation are as follows:(1)A feature extraction algorithm based on residual deep NMF has been developed.The algorithm constructs deep feature representation by cascading multiple NMF layers.Reconstruction residual of NMF is passed layer by layer to reduce information loss.Meanwhile,passing residuals between layers can construct a feature hierarchy from coarse to fine.Furthermore,activation functions are applied between adjacent layers to enhance the ability of non-linear feature extraction.(2)A feature extraction algorithm based on residual deep PCA is proposed.The algorithm is based on a multi-layer structure so that both shallow and deep features can be taken into account.Within each layer,PCA is utilized for layer-wise feature extraction.The reconstruction residual can get rid of information loss.The layer-wise features are concatenated to form the final output feature.On hyperspectral image datasets,the above two algorithms are compared with other existing algorithms.The experimental results show that the feature extraction algorithms based on deep matrix decomposition can effectively improve the classification accuracy of hyperspectral images.