Illumination condition prediction from given outdoor images is challenging but widely useful. Most of the existing methods predict the illumination based on the inserted markers or the limitation on the environment, which makes them impossible to directly predict the illumination condition from the nature images. We propose a multi-cues based illumination prediction approach that can predict the illumination condition by the input images only. Our approach combines the prediction results from the illumination-related information of input images. The illumination-related information is obtained by the proposed classification algorithm. The classification algorithm detects the geometry and shadow of the image at first, and then classifies the pixels of illumination-related cues by the requirement of the prediction algorithm. All of the illumination-related cues are used to predict the possible illumination condition by suitable algorithms, and the the markov random field is used to obtain the most likely illumination condition from all of the possible condition. Results of our approach and the comparison demonstrate the efficiency of our proposed multi-cues based illumination prediction approach.