Cloud Classification of Ground-based Infrared Images Based on Log-Euclidean Distance

Cloud type recognition of ground-based infrared images is a challenging task. A novel cloud classification algorithm on matrix manifolds is proposed to extract features of the images and to group them into five cloud types. The proposed algorithm comprises three stages: pre-processing, feature extraction and classification. Cloud classification is conducted by the Support Vector Machines (SVM) based on Log-Euclidean distance. The datasets are gathered by the whole sky infrared cloud measuring system and are divided into the standard and the actual-observed parts. The proposed method is compared to Calbo's work to verify its effectiveness. Seven dimensional features of the infrared cloud image are chosen for classification, including mean, standard deviation, smoothness, third moment, uniformity, entropy and correlation with clear, as recommended in Calbo's method. In the experiments, 50%, 60%, 70% and 80% of each class samples of the two datasets are selected randomly as training sets and the rest are treated as testing sets, respectively. We obtain overall type-recognition rates of 96.38% for the standard dataset and 83.07% for actual-observed dataset, while the results provided by Calbo's method are 90.38% and 81.11%, which indicates that the proposed model achieves a competitive recognition rate on the ground-based infrared images.