Using satellite imagery for cumulus cloud detection has an important significance for preventing meteorological disasters. Support Vector Machine (SVM), which can seek the best compromise between the complexity of the model and the learning ability based on finite sample information, is expected to play a role in the cumulus cloud detection. However, the traditional SVM is very sensitive to the samples of noise and outlier, and doesn’t possess the skill of fuzzy treatment, which doesn’t meet the fuzzy and uneven characteristics of satellite imagery and the complex and diverse cloud patterns. In order to solve the problem of SVM, this paper introduces Fuzzy Support Vector Machine (FSVM) and defines the range-rate of the distances from the adjacent samples to the class center, based on the distribution characteristics of training samples. Then, on the basis of the range-rate, we weed out the possible noises and outliers of training set and overcome the shortcoming that the affinity FSVM is susceptible to noises and outliers at the time of calculating the radius of smallest hyper-sphere, so as to make the obtained membership better reflect the variance of different sample sets. The experimental results show that, for FY2D satellite imageries, extracting 8-d spectral features from different channels, compared with traditional SVM and affinity FSVM, the accuracies of cumulus cloud detection based on the proposed method increase respectively by about 2% and 1%. The proposed method owns stronger adaptability and noise robustness, and can make better effect on early warning disastrous weather such as thunderstorm.