To address the problem that Bag-of-Words model still has several drawbacks such as the scarcity of information in single local descriptor, large quantization error and lack of representation upon image features in image classification tasks, an image classification method based on multi-descriptor hierarchical feature learning is proposed. Combing scale invariant feature transform (SIFT) and kernel descriptors-shape (KDES-S) features, a hierarchical structure is used to reduce quantization error in encoding process, which extracts local features. After that, image features in each layer are normalized respectively, the liner combination of which is the final feature representation for linear support vector machine (SVM) classifier. Experiments are conducted on datasets Caltech-101, Caltech-256 and Scene-15, and experimental results show that the proposed method improves the classification accuracy significantly in comparison with other algorithms.