Hyperspectral Image Dimensionality Reduction And Classification Based On Subspace

Hyperspectral remote sensing is a new technology for earth observation in the 20 th century.It has high spectral resolution and provides abundant information for ground objects.Hyperspectral remote sensing is a new type of remote sensing method.It plays an extremely important role in various fields such as the military and civilian fields,and has become a frontier of science and technology in the field of remote sensing.However,due to its characteristics such as multiple bands and large data volume,it poses a great challenge for image classification and recognition.This makes the phenomenon of “dimensional disaster” prone to occur when processing hyperspectral images,which leads to a decrease in classification accuracy.At present,many scholars at home and abroad are committed to studying various problems and difficulties encountered in the classification of hyperspectral remote sensing,and have achieved certain results.In order to avoid the phenomenon of “dimension disaster”,it is necessary to reduce the data volume and redundancy under the condition that the classification and recognition rate of hyperspectral features is high.Band selection of hyperspectral images is a very complex optimization combination problem,requiring that the selected waveband has a large amount of information,and that each class has better separability.Based on the summarization of existing research results,this article has conducted in-depth research on dimension reduction of hyperspectral images.The main contents are as follows:1?Hyperspectral image subspace division.Using the correlation between hyperspectral image bands,the entire spectral space of the hyperspectral data is divided into five initial classification intervals by automatic subspace division method according to the “blocking” characteristics of the band and correlation coefficient matrix images,and then In the initial classification interval,the clustering center of each initial classification interval is determined by the improved clustering algorithm,and the mutual information between the two adjacent cluster centers and their respective bands is calculated,and the band where the absolute value of the difference in mutual information is the smallest is found.Which is the boundary to divide subspaces.That is,the automatic subspace division method provides an initial classification interval for the improved clustering algorithm.The number of classification intervals is the number of cluster centers,and the mutual information is used to determine the boundary division subspace in the initial classification interval.2?Hyperspectral image the best band selection.In the divided subspace,the adaptive band selection method is used to solve the index values of all the bands,and the largest index is selected in each subspace.Five subspaces that are five indexes,and then the largest top 3 is selected among the five indexes.Indices,the bands corresponding to the three indices are the best combination of bands,based on this combination of bands,laying the foundation for subsequent classification of objects.3?Support vector machine classification.The specific process of object oriented image classification is given,and the classification principle of support vector machine(SVM)is introduced in detail.The optimal band combination is classified by SVM algorithm with radial basis function(RBF)as kernel function,and its overall classification accuracy,average precision and Kappa coefficient are obtained,and compared with the SVM classification results of the best band combination obtained by the Principal Component Analysis(PCA)reduction dimension,the Minimum Noise Fraction(MNF)and the Optimum IndexFactor(OIF)reduction.The results show that the overall classification accuracy of the two types of data using this method based on subspace best band selection is 87.43% and 85.92%,which the classification accuracy is higher than PCA,MNF and OIF dimensionality methods,which proves the correctness of the proposed method.