The Impact of Non-uniform Label Noise on the Classification of Hyperspectral Images

Abstract

Erroneous labels affect the learning models in supervised classification, deteriorate the classification performance and hinder thereby subsequent tasks. These erroneous labels are referred to as label noise. The influence of label noise on the classification performance has been so far mainly studied using simulations with a uniform distribution of noisy labels accross the classes. In this paper, we propose a new label noise simulation approach for hyperspectral images (HSIs) where similar classes have a higher chance to be mixed up. Under such a realistic label noise simulation model, we compare the behaviour of different types of classifiers for HSI in remote sensing, including traditional machine learning and deep learning approaches. Our analysis reveals which levels of label noise are acceptable for a given tolerance in the classification accuracy and how robust are different learning models in this respect.