Identifying factors impacting the overall accuracy in image classification problems: a statistical approach

Abstract

Image classification is a subject of pattern recognition that can be applied in several areas. Obtaining highly-accurate classification involves choosing optimal set-ups from which images will be classified. In this process, controllable variables can affect the overall classification accuracy, such as the image’s spatial resolution and the classification method. In this sense, we have designed a factorial experiment where the classification accuracy of an image (from Curitiba, Paraná, Brazil) was obtained from three satellites and three classification methods. The Kruskal-Wallis test was applied to evaluate if the variability across factor levels supports the hypothesis that the experimental factors’ effects are statistically significant. Then, we evaluated which factor levels differed from each other using post- hoc tests. Our findings suggest that the image’s spatial resolution and the interaction between Satellite and Classification Method are determinants in obtaining accurate image classifications in a geographical context.