This paper describes advances in an automatic approach for the detection of sand dunes of Mars, based on supervised learning techniques. A set of features (gradient histogram) is extracted from the remotely sensed images and two classifiers (Support Vector Machine and Random Forests) are trained from this data. The evaluation is conducted on 230 MOC-NA images (spatial resolution between 1·45 and 6·80 m/pixel) leading to about 89% of correct detections. A detailed analysis of the detection results (dune/non-dune) is performed by dune type or bulk shape, confirming high performances independently of the way the dataset is analysed. This demonstrates the robustness and adequacy of the automated approach to deal with the large variety of aeolian structures present on the surface of Mars. Copyright © 2012 John Wiley & Sons, Ltd.