Hyperspectral Band Selection Based On Representation Learning And Fuzzy Clustering

Hyperspectral remote sensing is the frontier research field of remote sensing.Hyperspectral images(HSIs)acquired by imaging spectrometers contain rich spectral and spatial information of target objects,making HSIs widely used in many fields such as environmental monitoring,agricultural disaster assessment,and ocean observation.However,the rich spectral information in HSIs also leads to the problems of high dimensionality and excessive redundant information,which pose challenges to applications such as HSIs classification.Therefore,it is necessary to perform dimensionality reduction on HSIs.Band selection is an important method for the dimensionality reduction of HSIs,which aims to select a set of representative bands from HSIs with a large amount of information,good category separability,and low correlation to represent the original dataset.Band selection can effectively preserve the original features of HSIs during dimensionality reduction and it has become a current research hotspot.In recent years,clustering-based band selection methods have recently demonstrated a promising prospect for reducing information redundancy and improving the classification performance of HSIs.However,most existing clustering-based band selection methods conduct clustering on original HSIs,which limits their performance because of the high dimensionality of hyperspectral bands.To tackle this problem,it is a new idea to combine representation learning and clustering for band selection.Therefore,designing an effective joint model of clustering and representative learning for the band selection task and introducing appropriate regularizations into the joint model based on problem-dependent information remains challenging.To address the above problems,this paper proposes the HSIs band selection method based on joint learning of representation learning and fuzzy clustering.The main research contents are as follows:1.A novel hyperspectral band selection method termed joint learning of correlationconstrained fuzzy clustering and discriminative nonnegative representation for hyperspectral band selection(CFNR)is presented in this paper.To tackle the problem of most existing clustering-based band selection methods conducting clustering on original HSIs,which limits their performance,CFNR introduces graph regularized nonnegative matrix factorization into the fuzzy C-means(FCM)model.This can implement the joint learning of nonnegative feature representations of bands and fuzzy clustering and exploit the intrinsic manifold structure of HSIs to learn low-dimensional discriminative nonnegative representations of bands for better clustering.In addition,a correlation constraint is designed to enforce the similarity of clustering results between neighboring bands,which is imposed on the membership matrix of FCM in the proposed method to obtain clustering results that meet the needs of band selection.2.A novel hyperspectral band selection method termed joint learning of multi-graph regularization and low-rank constrained nonnegative matrix factorization and fuzzy clustering(MLNFC)is proposed.The CFNR model lacks consideration of the global correlation among all bands in HSI in feature representation learning.To tackle this problem,MLNFC introduces a global low-rank constraint for nonnegative representation learning,which can effectively utilize the global correlation information of the HSI.In addition,the proposed method introduces multi-graph regularization into the proposed method,which can capture a more accurate intrinsic manifold structure between bands.This also can obtain clustering results that are beneficial to band selection on a more effective low-dimensional feature representation.The models in this paper are solved by the alternating direction multiplier method.At the same time,to verify the effectiveness of the proposed method,three evaluation indexes,Overall Accuracy,Average Overall Accuracy,and Kappa coefficient,are used for experimental comparison with several state-of-the-art methods on three real hyperspectral datasets.Experimental results show the effectiveness of the method.