Performance Evaluation of Machine Learning Classifiers for Hyperspectral Images

Abstract

Hyperspectral remote sensing is a popular research area that has gained substantial interest in recent years. Hyperspectral images are characterized in hundreds of spectral bands with high spectral resolution and this immense information reflects many features. This paper focuses mainly on range of Machine Learning (ML) algorithms for classification of hyperspectral images using Principal Component Analysis (PCA) as dimension reduction technique. Popular ML classifiers such as Random Forest (RF), Logistic Regression (LR), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Classification and Regression Trees (CART), Naive Bayes (NB), Stochastic Gradient Descent (SGD) Classifier, XGBoost and Multilayer Perception (MLP) are applied for 8 publicly available datasets. In this paper model performance evaluated for different Principal Components(PC) considering total model accuracy. In comparison with other classifiers MLP which is based on Neural Network (NN) performed well for all the selected datasets. However RF, XGBoost & KNN also shows promising performance. Evaluation for the first 15 PC with MLP, Pavia Centre(PaviaC) dataset achieved 0.99 model accuracy whereas Salinas Corrected(SalinasC), Salinas, Pavia University(PaviaU), Kennedy Space Center(KSC), Indian Pines Corrected(IPC), Indian Pines(IP) and Bostwana achieved 0.96, 0.94, 0.96, 0.95, 0.83, 0.79, 0.96 total model accuracy respectively.