Pattern classification methods have been widely studied for analysis of brain images to decode the disease states, such as diagnosis of Alzheimer's disease (AD). Most existing methods aimed to extract discriminative features from neuroimaging data and then build a supervised classifier for classification. However, due to the rich imaging features and small sample size of neuroimaging data, it is still challenging to make use of features to achieve good classification performance. In this paper, we propose a hierarchical ensemble classification algorithm to gradually combine the features and decisions into a unified model for more accurate classification. Specifically, a number of low-level classifiers are first built to transform the rich imaging and correlation-context features of brain image into more compact high-level features with supervised learning. Then, multiple high-level classifiers are generated, with each evaluating the high-level features of different brain regions. Finally, all high-level classifiers are combined to make final decision. Our method is evaluated using MR brain images on 427 subjects (including 198 AD patients and 229 normal controls) from Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Experimental results show that our method achieves an accuracy of 92.04% and an AUC (area under the ROC curve) of 0.9518 for AD classification, demonstrating very promising classification performance.