Spectral–Spatial Classification of Hyperspectral Images Using Probabilistic Weighted Strategy for Multifeature Fusion

Abstract

A hyperspectral image classification method based on probabilistic weighted fusion of multiple spectral–spatial features is proposed in this letter. First, dimensionality reduction and feature extraction of hyperspectral images are conducted by minimum noise fraction. Then, two spectral–spatial features are composed through a combination of texture features and multiscale morphological features with characteristic images obtained by minimum noise fraction. Next, a support vector machine classifier is employed to classify each spectral–spatial feature. Finally, a probabilistic weighted fusion model is established and applied for probabilistic fusion of the classification output of every single feature. The proposed method is validated by classification experiments using Reflective Optics Spectrographic Imaging System (ROSIS) and Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) images, which achieves the overall accuracy of 96.32% and 96.49%, respectively. The results indicate that the accuracy of hyperspectral image classification is improved effectively by the proposed method and the method has superior classification performance to other conventional multifeature fusion methods.