Unsupervised band removal leading to improved classification accuracy of hyperspectral images
ACSC '06 Proceedings of the 29th Australasian Computer Science Conference - Volume 48
Blind spectral unmixing by local maximization of non-Gaussianity
Signal Processing
Novel classification and segmentation techniques with application to remotely sensed images
Transactions on rough sets VII
Unsupervised classification of hyperspectral images on spherical manifolds
ICDM'11 Proceedings of the 11th international conference on Advances in data mining: applications and theoretical aspects
| Hi-index | 0.00 |
Conventional remote sensing classificationtechniques that model the data in each class with amultivariate Gaussian distribution are inefficient, as thisassumption is generally not valid in practice. We presenta novel, Independent Component Analysis (ICA) basedapproach for unsupervised classification of hyperspectralimagery. ICA, employed for a mixture model, estimatesthe data density in each class and models classdistributions with non-Gaussian structure, formulatingthe ICA mixture model (ICAMM).We apply the ICAMM for unsupervised classificationof a test image from the AVIRIS sensor. Four featureextraction techniques namely Principal ComponentAnalysis, Segmented Principal Component Analysis,Orthogonal Subspace Projection and Projection Pursuithave been considered as preprocessing steps for reducingthe data dimensionality. The results demonstrate that theICAMM significantly outperforms the K-means algorithmfor land cover classification of hyperspectral imageryimplemented on reduced data sets. Moreover, datasetsextracted using Segmented Principal Component Analysisproduce the highest classification accuracy.