Segmented Nonnegative Matrix Factorization for Hyperspectral Image Classification

The remote sensing hyperspectral image (HSI) consists of hundreds of narrow and adjoining spectral bands. It carries a lot of significant information about the earth's objects. However, the use of all HSI bands leads to higher misclassification. Band reduction is a potential solution to resolve this issue, where feature selection and feature extraction methods are commonly accomplished for the reduction of bands. One of the most commonly used unsupervised feature extraction techniques is the Principal Component Analysis (PCA). But it fails to bring out the local intrinsic information from the HSI as it ponders only the global variation of the data. This problem can be addressed by the Segmented PCA (SPCA) which exploits both the global and local variance of the data by partitioning it into highly correlated blocks. Beside, another unsupervised feature extraction technique named Nonnegative Matrix Factorization (NMF) is also applied for HSI by approximating the data in a low-dimensional subspace. In this paper, we propose a feature extraction method, named Segmented Nonnegative Matrix Factorization (SNMF), performing NMF on the segmented strongly correlated blocks of HSI data. The efficacy of the proposed method is compared with PCA, NMF, and SPCA on the Indian Pines dataset with a support vector machine classifier. The experimental result shows that SNMF (89.00%) outperforms PCA (84.33%), NMF (85.37%), and SPCA (87.59%) over all classes' samples.