A cumulonimbus detection approach was proposed based on Multi Fuzzy Support Vector Machine(FSVM) by using a decision fusion strategy to solve the contradiction that adding more features will increase the accuracy of cloud classification while cause over fitting phenomenon due to high feature dimensions. Firstly, spectral features, the brightness temperature difference of multi-channels, first order histogram texture features, gray level co-occurrence matrix texture features and Gabor wavelet features were extracted from training cloud images to form a training sample set which contains 5 kinds of features. Then, five FSVM sub-classifiers were trained respect to each kind of feature. Finally, the output of each sub-classifier was fused by weighted decision in the output space to improve the detection accuracy of the cumulonimbus. Experimental results show that the proposed approach solves the over fitting phenomenon in cumulonimbus detection caused by the too high feature dimensions and can determine the weight of different features adaptively. The results also demonstrate that the accuracy is not only superior to each FSVM sub-classifier but also to the FSVM classifier trained by all the input features at once. The proposed approach is expected to be applied in the analysis of satellite cloud images.