Ground-Based Cloud Images Classification Based On SVM-Classification-Tree

Nowadays, Clouds are playing an essential role in the atmosphere and cloud classescan be seen as indicators of the type and intensity of locally active thermal processes.Cloud cover, type and height measurements are collected by means of both satellite andground-based weather stations. However, the satellite images data storage space is so vastthat it can’t afford a real-time process operation. Consequently, we have to predict weatherin a short time and local area by using the ground-based cloud images,In this paper, we propose a novel method for ground-based cloud classification. Wefirstly compute different texture features from the ground-based cloud images according tothe images’ characteristics, and we select three feature groups from lots of features forimage analyze. But with further experimentations we can find out that features fromGabor feature group have high discrimination performance. The Gabor feature group isbased on2D Gabor filters with the theory of multichannel decomposition. These featuresare more robust and effective than the well-known features such as gray-levelco-occurrence matrix (GLCM), the Run-Length Texture Feature and the structure features.Due to the training samples are usually selected randomly, which can cause highcomplexity of the classifier, thus the training samples were clustered into several clustersby using the supervised clustering algorithm. And a small quantity of representativeinstances was chosen as training sets.At last, Ensemble SVMs and binary-tree SVMs were constructed. According to lotsof experimentations, the result shows that these methods have higher classificationaccuracy and lower space complexity comparing to the other classification algorithms.