Evaluation of supervised classification of hyperspectral imagery

Abstract

Hyperspectral Imaging (HSI) is a powerful source of information for land cover and land use classification maps. HSI sensors trade off spatial resolution to gain high spectral resolution, measuring hundreds of electromagnetic bands at once. Variations in spectrum enable classifiers to identify the materials and composition of the earth's surface. This work will evaluate current methods of classifying hyperspectral imagery using supervised learning. In supervised learning, the classifier receives a training set of pixels assigned class labels. The training set is developed by using ground truth, or information known about the earth's surface derived from ground surveys and other sources. Since this process is both expensive and time consuming, favorable classifiers achieve high accuracy from a small training set. The classifier learns from the training set how to assign class labels to additional, unlabeled pixels. The final classification map is an image whose pixel values map a class label to a point on the ground. The goal of this work is to holistically examine the process of classifying HSI using supervised learning, from creating a training set and training the classifier to evaluating the accuracy of the classification map. We approach this process from the viewpoint of an analyst with a critical eye to the realistic limitations in how to collect ground truth and build the training set. Our analysis seeks to give an unbiased assessment of supervised classifier performance and to examine the generality and applicability of classifiers to real-world problems. To accomplish this, we analyzed several common types of classifiers through an exhaustive experimentation regime to identify holes or biases in classifier performance. During testing, we varied four key parameters, including the dataset, classifier, sampling mode, and amount of labeled data used to train the classifier. Due to the high cost of labeling data, we used six HSI datasets common to academic research. Our analysis yielded interesting conclusions about the methodology of analysts classifying a HSI datacube and considerations for transforming current research into tools for analysts. We found that the basic linear discriminant analysis and quadratic decision analysis classifiers matched or exceeded more complex neural networks with a fraction of the training time under certain conditions. Our recommendations for the application of supervised learning to hyperspectral image classification include the minimum number of samples per class for training, the method to select samples for the training set, and when to employ simple versus complex classifiers on a scene.--Author's abstract