Sensitivity of hyperspectral classification algorithms to training sample size

Abstract

Algorithms that exploit hyperspectral imagery often encounter problems related to the high dimensionality of the data, particularly when the amount of training data is limited. Recently, two algorithms were proposed to alleviate the small sample size problem - one is based on employing a Multi-Classifier Decision Fusion (MCDF) in the raw reflectance domain, and the other employed the MCDF framework in the Discrete Wavelet Transform domain (DWT-MCDF). This paper investigates the sensitivity of conventional single classifier based classification approaches, as well as MCDF and DWT-MCDF to variations in the amount of data employed for training the classification system. The hyperspectral data in this experiment was obtained using an airborne hyperspectral imager used by SpecTIR™. The results of the experimental analysis show that for the given application, the MCDF and DWT-MCDF algorithms are significantly less sensitive than the conventional algorithms to limited training data. PCA consistently results in overall accuracies of about 35%. LDA accuracies are very high, about 75%, when there is an abundance of training data - about 10X (i.e. number of training samples is 10 times the number of spectral bands); remains above 60% for training data abundances of 2X and higher; but dramatically decreases to ∼20% for abundances of 1X. MCDF results in accuracies ranging between 65% and 75% for training data abundance of 3X and higher, but the accuracies drop to ∼60% for 2X and ∼55% for 1X. DWT-MCDF results in high accuracies with the least sensitivity to training data abundance. Its accuracies range between ∼60–65% for abundances of 1X to 10X.