The paper proposes a new classification feature: DI(Diversity Index), considering the characteristics of satellite cloud image. The DI feature presents the structure of cloud effectively and it has a good robustness. The DI feature avoids the influence exerted by the variety of cloud positions. This paper proposes the semi-supervised FCM (SSFCM) method in the domain of satellite cloud images classification. The SSFCM method overcomes the blindness brought by the FCM method without considering the domain knowledge. The SSFCM method uses a small number of samples labeled by experts to direct the clustering process through comparing with the labeled samples in terms of similarity. These labeled samples represent the domain knowledge. The experiments demonstrate that the SSFCM method improves the accuracy of cloud classification based on the DI feature.