To improve the accuracy in scene categorization of satellite images,this paper presents an algorithm of feature selection based on augmented LDA(Latent Dirichlet Allocation) model,and the algorithm is improved,which can automatically selects features from the features-pool.This method firstly extracts five kinds of features(SIFT,Geometric Blur,LBP,Gabor and Color histogram) from each image,and during the cross-validation,the combined features,which have the best performance over the dataset are got.Next,the dimensionality of the combined features is reduced by using LDA.Finally the regularized logistic regression classifier are employed to achieve the classification.Compared with other feature combination,the experimental results demonstrate that,the combination of the automatically selected features can improve the accuracy of scene categorization of satellite images effectively.