Research And Application Of Hyperspectral Object Classification Based On Discriminative Dictionary Learning And Sparse Representation

Hyperspectral images(HIS)classification is a major research content in remote sensing field,and it is also an important method to excavate feature-related information from remote sensing images,which has been widely applied in military and civil fields.While hyperspectral images contain abundant spectral and spatial information,they also bring challenges such as high pixel features,multiple information redundancy and low classification efficiency.Thus,sparse and low-rank representation and dictionary learning at pixelwise HIS classification have been more mature and valid.In addition,the introduction of spatial information also significantly improves the classification effect.The innovation work and main tasks of this paper are as follows:(1)Low-rank representation based discriminant sub-dictionary learning algorithm was proposed.The proposed method separately learns a sub-dictionary for each class and introducing global constraints in the training process.Then,perform low-rank reconstruction on HIS to remove the influence of noise while retaining the global features.In the classification stage,multi-scale neighbors strategy merges all pixels in the neighbors to classify.Experiments demonstrate that compared with the traditional algorithms,the proposed algorithm obtained a significant improvement in quantitative and qualitative analysis.It is also robust against imbalanced datasets.(2)Sparse representation based adaptive multi-scale superpixel classification algorithm was proposed.First,PCA is performed on the HIS and first three principal components are selected and merged into a three-channel color image.Then,the color image multi-scale superpixel segmentation is performed.Finally,the superpixel segmentation result is mapped to the original dataset,and all pixels from superpixel are merged for sparse reconstruction.The proposed algorithm can adaptively change the shape and size of superpixel according to the type of feature structure and can incorporate more and more accurate spatial information into the classification process.Experiments verify the effectiveness of the proposed algorithm,broaden the application scenarios of dictionary learning algorithms,and improve the algorithm's robustness to different types of datasets.(3)Designed and implemented a hyperspectral image classification system.The system includes four functional modules: data and label reading module,dimensionality reduction and superpixel segmentation module,dictionary training module and classification test module.The system is a practical application of the first two algorithms,which can clearly show the algorithm process and classification effect and facilitate comparative analysis with existing image classification algorithms.