Adaptive Fuzzy Support Vector Machine for Classification of Clouds in Satellite Imagery

The classification of clouds plays an important role in analyzing satellite imagery automatically. Specific to the characteristics that satellite imagery is susceptible to noises and different types of clouds tend to overlap, a classifier based on adaptive fuzzy support vector machine (AFSVM) for classification of clouds is constructed. First, the classifier confirms a minimum hypersphere to distinguish effective samples and no-effective samples by Support Vector Data Description (SVDD) method. The samples inside of the hypersphere are taken as effective samples, while the samples outside of the hypersphere are regard as no-effective samples. Then the formula of membership function is modified to make that the membership attenuation speed of the effective samples is slower than that of the no-effective samples. Finally, the paper defines three parameters to control specially the critical membership degree and the attenuation trend of membership function. The proposed membership function overcomes the shortcoming of the traditional FSVM that its membership function could not describe the distribution of samples effectively, and makes itself adjust adaptively according to the specific distribution characteristics of different cloud sample sets. Experiments were conducted on MTSAT satellite imagery, the results showed that by extracting the spectral feature of the albedo of VIS channel, the bright temperatures of four IR channels and the three bright temperature differences as the spectral features, and combined with the statistical texture features, the proposed classifier is able to distinguish the clear weather, low clouds, middle clouds, high clouds and clouds with vertical development effectively with high accuracies, and the performances are superior to the standard SVM and traditional FSVM in terms of stability and adaptability.