A hybrid method based on extreme learning machine and k-nearest neighbor for cloud classification of ground-based visible cloud image

Abstract

A classification scheme based on extreme learning machine and k nearest neighbor is proposed for cloud classification. In this work, 21 characteristic parameters of texture features, color features and shape features are selected from four different sky conditions (cumulus, stratus, cirrus and clear sky) for classification. The results show that the new scheme using texture features, color features and shape features together can get better performance than using these features alone or any two of them together. When all 21 features are used for classification, the accurate identification rates of cumulus, stratus, cirrus and clear sky are 84.56%, 78.06%, 76.67% and 100.00% respectively, with an average of 84.82%. The proposed model can benefit from the merits of the k-nearest neighbor and the extreme learning machine through its novel structure with high robustness particularly for cloud classification. The simulation results demonstrate that the proposed model in this work is practical for cloud classification and outperforms extreme learning machine (ELM) models, artificial neural network (ANN), k-nearest neighbor (KNN), hybrid method based on KNN and ANN (KNN–ANN), and support vector machine (SVM).