This paper focuses on automatic road extraction in urban areas from high resolution satellite images. We propose a new approach based on machine learning. First, many features reflecting road characteristics are extracted, which consist of the ratio of bright regions, the direction consistency of edges and local binary patterns. Then these features are input into a learning container, and AdaBoost is adopted to train classifiers and select most effective features. Finally, roads are detected with a sliding window by using the learning results and validated by combining the road connectivity. Experimental results on real Quickbird images demonstrate the effectiveness and robustness of the proposed method.