Aurora Image Classification Robust to Noise Labels

In recent years, a lot of researcher has analyzed a shape of aurora from aurora images using machine learning. In particular, a number of studies have shown that convolutional neural network (CNN) can be used to classify aurora images with high accuracy. We analyzed aurora images and found that there is a certain amount of mislabeled data in aurora datasets. As this labeling error causes the difficulty in judging the shape of aurora, it is important for aurora image classification to assign accurate labels to aurora images. However, the shape of aurora is ambiguous and it is very difficult to label the aurora image accurately. We think that these mislabeled data are hindering aurora image analysis. In this study, we propose a classification method and a CNN training method that reduces the effect of mislabeled data. Specifically, we used Nearest Neighbor Search (NNS) for aurora image classification instead of a final layer of a pre-trained CNN model. In addition, we propose a simple method for determining mislabeled data. Using a combination of these two methods to automatically classify aurora images, we achieved an F1 score of 0.945 for an aurora image dataset with four labels: arc aurora, drapery, hot-spot and radial, an improvement of 0.2 over traditional CNN-based classification.