The need for computers to make educated decisions is growing. Various methods have been developed for decision making using observation vectors. Among these are supervised and unsupervised classifiers. Recently, there has been increased attention to ensemble learning--methods that generate many classifiers and aggregate their results. Breiman (2001) proposed Random Forests for classification and clustering. The Random Forest algorithm is ensemble learning using the decision tree principle. Input vectors are used to grow decision trees and build a forest. A classification decision is reached by sending an unknown input vector down each tree in the forest and taking the majority vote among all trees. The main focus of this research is to evaluate the effectiveness of Random Forest in classifying pixels in multispectral image data acquired using satellites. In this paper the effectiveness and accuracy of Random Forest, neural networks, support vector machines, and nearest neighbor classifiers are assessed by classifying multispectral images and comparing each classifier's results. As unsupervised classifiers are also widely used, this research compares the accuracy of an unsupervised Random Forest classifier with the Mahalanobis distance classifier, maximum likelihood classifier, and minimum distance classifier with respect to multispectral satellite data.
Date of publication
Lowe, Barrett E., "The Random Forest Algorithm with Application to Multispectral Image Analysis" (2015). Computer Science Theses. Paper 5.