A Dimensionality Reduction Method Based On Frequency-domain Similarity And Spatial Structure Preservation

Dimensionality reduction(band selection)is a data processing technique used to simplify hyperspectral image data and extract important features,reducing data redundancy and noise,which can improve data processing efficiency while retaining the main information in the original data,facilitating subsequent analysis and applications.In this paper,we first provide theoretical proof for the ADMM(Alternating Direction Method of Multipliers)algorithm to ensure its rationality for subsequent applications.Secondly,considering the piecewise smoothness of spectral bands and the spatial structure relationship of spectra,we construct a dimensionality reduction matrix and optimize it using the ADMM algorithm,named SVM-FS algorithm,to maximize the retention of original information and spectral structure while considering both frequency and spatial domains for band selection.The dimensionality reduction method’s efficacy was confirmed by evaluating its classification performance on four authentic datasets:Indian Pine,Pavia University,Salinas,and Kennedy Space Center.