Due to the problem that traditional latent semantic analysis (LSA) method is unable to obtain spatial distribution information of objects and discriminative information of latent topic. We propose a scene classification approach based on multi-scale spatial discriminative probabilistic latent semantic analysis (PLSA). First

it decomposes images in multiple scales using a spatial pyramid approach to obtain spatial distribution information for images. Then

the PLSA model is used to extract the latent semantic information of each local block. Next

the latent semantic features of all local blocks are concatenated with different weights to produce the multi-scale spatial latent semantic information of image. Finally

we exploit weight learning method to learn the discriminative information between different image topics and get multi-scale spatial discriminative latent semantic information of image. Afterwards

the weight information is integrated into the support vector machine (SVM) classifier to perform image classification. Experimental results on the common three scene image datasets

viz. Scene-13

Scene-15 and Caltech-101

demonstrate that our method performs much better than the existing state-of-the-art approaches. Which demonstrate the importance of spatial information and discriminative information in image classification and further verify the effectiveness and robustness of our approach.