High performance of the support vector machine in classifying hyperspectral data using a limited dataset

摘要

To prospect mineral deposits at regional scale, recognition and classification of hydrothermal alteration zones using remote sensing data is a popular strategy. Due to the large number of spectral bands, classification of the hyperspectral data may be negatively affected by the Hughes phenomenon. A practical way to handle the Hughes problem is preparing a lot of training samples until the size of the training set is adequate and comparable with the number of the spectral bands. In order to gather adequate ground truth instances as training samples, a time-consuming and costly ground survey operation is needed. In this situation that preparing enough field samples is not an easy task, using an appropriate classifier which can properly work with a limited training dataset is highly desirable. Among the supervised classification methods, the Support Vector Machine is known as a promising classifier that can produce acceptable results even with limited training data. Here, this capability is evaluated when the SVM is used to classify the alteration zones of Darrehzar district. For this purpose, only 12 sampled instances from the study area are utilized to classify Hyperion hyperspectral data with 165 useable spectral bands. Results demonstrate that if parameters of the SVM, namely C and σ, are accurately adjusted, the SVM can be successfully used to identify alteration zones when field data samples are not available enough.