كاهش بعد تصاوير فراطيفي از طريق خوشه بندي فازي باندها

Dimensionality Reduction of Hyperspectral Images via Bands Fuzzy Clustering

This paper proposes an innovative band selection method based on fuzzy clustering of bands for hyperspectral images. The main novelties of this research lie in two issues: i) bands representation in a new space called prototype space (PS) where bands take characteristic vector in terms of class reflectivity (where bandsʹ feature vectors are defined in terms of class reflectivity); and ii) using uncertainty and angle measures to distinguish highly correlated and informative channels. Having clustered channels by Fuzzy C-Means (FCM) clustering in PS, the highly correlated bands fall in a cluster by uncertainty measure; then a nearest band to the cluster center is discerned as representative of those bands. Moreover uncertain bands as isolated bands are separated in PS where among isolated bands the bands that have large angle with diagonal of PS are indicated as informative channels. Since finding representative and informative bands depends on random initialization of FCM clustering, the bands clustering algorithm conducted many times. Accordingly, the proper bands are selected by maximizing overall accuracy of validation data set. Benchmarking on the challenging hyperspectral data set demonstrated relative merit of proposed method respect to the conventional band selection methods like sequential forward floating and sequential backward floating band selection.