Tensor-Based and Projection-Based Methods for Dimensionality Reduction of Hyperspectral Images

Abstract

Inasmuch as the hyperspectral images are represented by large amounts of data it is necessary to adopt an appropriate method to reduce their size without affecting the quality of their processing results. This paper addresses the use of Tucker1 decomposition for tensor compression and dimensionality reduction, followed by a projection-based method, Principal Component Analysis (PCA). After dimensional reduction, some classifications were performed using the features extracted. Various supervised learning algorithms were used for which we calculated k-fold cross-validation loss. We made a comparison of these methods in terms of the classification results obtained. According to the results, at the same size of the transformed data, PCA features have led to lower accuracy than Tucker1 ones, and the original data.