Effective Feature Extraction via Segmented t-Stochastic Neighbor Embedding for Hyperspectral Image Classification

Abstract

Remote sensing hyperspectral imagery (HSI) contains significant information about the earth’s objects captured by hundreds of narrow and adjoining spectral bands. Considering all the bands for classification, the performance gets hampered. Consequently, it is crucial to reduce the HSI bands usually via feature extraction and feature selection. Principal Component Analysis (PCA) is one of the broadly used unsupervised feature extraction techniques. However, it considers global variance neglecting the local structure of the data. Segmented PCA (SPCA) overcomes this problem of PCA to some extent. Although t-Stochastic Neighbor Embedding (t-SNE), another unsupervised feature extraction technique, can be effectively used for data visualization and feature extraction, T-SNE is a probabilistic approach that tries to preserve the local structure of the dataset by persevering the relative distance between the data points. More local subtle characteristics can be considered by performing segmentation before applying t-SNE like SPCA. As such, in this paper, we propose the Segmented t-SNE (Seg t-SNE) feature extraction method exploiting the benefits of segmentation and t-SNE together. To analyze the efficacy, the performance of our proposed method Seg t-SNE is compared with PCA, SPCA, and t-SNE. The experimental outcomes demonstrate that Seg t-SNE (88.70%) outperforms PCA (84.39%), all the variants of SPCA (maximum of 88.59%), and t-SNE (77.52%) considering all classes’ samples.