Scenario Analysis for Image Classification using Multi-objective Optimization

Abstract

In a typical image classification task, the analyst decides beforehand the number of classes and which image channels to use. If there is a need to modify the classes or data channels, it is necessary to start over. This paper proposes a scenario analysis tool for the task of image classification as a way of automating this process. Each scenario represents the parameters that will be used in a complete supervised classification task, including training and classification. The proposed method uses multi-objective optimization to evaluate different sets of attributes and classes, and presents the compromising solutions, regarding the user objectives. A class hierarchy structure is used to generate different class sets, and the system attempts to find the most appropriate combinations of class and attribute sets. In this work, the system is applied to remote sensing problems and we consider three objectives: the best classification accuracy, the smallest attribute set and the biggest class set. The system shows the compromising combinations of class and attribute sets, along with the accuracy on a testing sample. The user can then choose which combination to use for the image classification.