Research On Hyperspectral Image Classification Algorithm Based On Pixel Information And Collaborative Representation

Hyperspectral images play a vital role in forestry inspection,urban planning,product quality inspection,mineralogy and other fields.Compared with traditional two-dimensional images that only contain spatial information,hyperspectral image classification can obtain more detailed spectral information,which can effectively improve the ability to recognize objects and improve the classification accuracy.Therefore,the effective use of spatial information and spectral information can be achieved through the feature extraction method.However,hyperspectral images have the problems of high band correlation,complex spatial structure,and high dimensionality of data,which restrict the improvement of classification ability.In this thesis,starting from feature extraction and local information of pixels,the following two improved classification algorithms are proposed.Hyperspectral images have high similarity between spectral bands and there are a large number of high-dimensional nonlinear samples.Traditional representation-based classification methods cannot effectively distinguish different samples in the same band and will cause dimensional disaster,which ultimately affects the classification performance.To this end,a hyperspectral classification algorithm based on spatial spectrum fusion and collaborative representation is proposed.A discriminative feature dictionary is constructed by alternately learning spatial and spectral features and used for spatially aware collaborative representation.During the classification process,the correlation coefficient between the feature dictionary and the test sample is calculated,and it is fused with the error decision.Experiments are carried out on two hyperspectral image datasets,and the results show that the experimental results of the method are better than the control experiments,which verifies the effectiveness of the algorithm in this thesis.In the process of hyperspectral image classification,the traditional sparse representation algorithm pays too much attention to the sparseness of the dictionary,but does not take into account the local spatial constraint information,which causes the algorithm to not fully utilize the spatial information,and the sufficient spatial information Expression is crucial.To this end,a sparse representation algorithm of spatial fusion and local location constraint dictionary is proposed.First,the spatial location information is introduced into the sample.Then,use the K-neighbor algorithm to calculate the Euclidean distance between the test sample and the training sample and the top K indices with the smallest distance in the local space,and construct a feature dictionary with both local constraints and spatial location information.Finally,according to the principle of minimum error,the predicted labels are assigned to the test samples to complete the final classification.In the comparison with the sparse representation algorithm of local constraints,it can be seen that the introduction of spatial location information can indeed improve the discriminativeness of the dictionary,thus improving the classification ability.