Publication

Abstract : Dimensionality reduction is an important step in processing hyperspectral images (HSIs) to overcome the dimensionality problem. Linear dimensionality reduction methods such as independent component analysis (ICA) and linear discriminant analysis (LDA) are commonly employed to reduce the dimensionality of HSI. These methods fail to capture nonlinear dependency in the HSI data, as data lies in the nonlinear manifold. To handle this, nonlinear transformation techniques based on kernel methods were introduced for dimensionality reduction of HSI. However, the kernel methods involve cubic computational complexity while computing the kernel matrix, thus its potential cannot be explored when the number of pixels (samples) is large. In literature, a fewer number of pixels are randomly selected to partially overcome this issue; however, this suboptimal strategy might neglect important information in the HSI. We propose randomized solutions to the ICA and LDA dimensionality reduction methods using random Fourier features, and we call them randomized ICA and randomized LDA. Our proposed method overcomes the scalability issue and handles the nonlinearities present in the data more efficiently. Experiments conducted with three real-world hyperspectral datasets demonstrate that our proposed randomized methods outperform the conventional kernel ICA and kernel LDA overall and in terms of per-class accuracies, and computational time.