Abstract: Scene image classification has recently been popular. The classification methods based on statistical models have been the most important methods in scene classification. We propose a new scene image classification framework based on multi-feature and an extended pLSA model We extract multiresolution histogram moments features and scale invariant feature transform （SIFT） features of patches of images. These patches are extracted on regular segmentations of different scales of every image. Both the features are scale invariant, so they can well describe the characteristic of image patches of different scales. At last, we use extended pLSA to model all training images. Test images are then dealt with a method called fold in. Our methods are not only unsupervised, but also can well represent semantic characteristic of images. We conduct three experiments on three often used image databases. We compare our methods with two previous baseline methods. And our methods get better results than the others