In this paper, we propose an aurora classification method using a multi-level feature representation aimed to capture both global and local texture information, and to reduce the feature space dimension substantially. First-order and second-order statistics are computed for an input image and its low-frequency scaled images at three lower levels obtained using wavelet decomposition. The features include gray level distribution, co-occurrence matrix features, and run-length matrix features. A support vector machine (SVM) classifier was trained and tested on a Chinese Arctic Yellow River Station dayside aurora image dataset. Classification performance was evaluated and compared with those of k-nearest neighbor (KNN) classifiers and back-propagation neural networks (BPNN). To explore the possibility of using a smaller feature space, we used a Minimum-Redundancy Max-Relevance feature selection strategy. The result shows that there is only indistinct performance decrease by reducing the feature vector from a total of 88 to the most discriminatory 38 features. This proves that our multi-level feature representation is very robust.