This paper proposes a new forest smoke detection method using spatial-temporal visual features extracted from camera images and a pattern classification technique. First, moving regions are detected by analyzing the frame difference between two consecutive key frames. Since smoke regions generally have a similar color, simple texture, and upward motion, the intensity, wavelet coefficients, and motion orientation are extracted as visual features. In addition, random forests are constructed using training data and then used for smoke verification process with four smoke classes. The proposed algorithm is successfully applied to various forest smoke videos and shows a better detection performance when compared with other methods.