This document presents a real-time image processing technique to detect steam in video images of an oil sands mining operation. The technique treats steam detection as a supervised pattern recognition problem. It uses a dual-tree complex wavelet transform to extract features from local regions of video frames. A statistical hidden Markov tree model characterizes the distribution of wavelet coefficients, capturing how texture patterns change with the presence of steam. These features are then classified using a support vector machine to detect steam-covered regions in each frame. The method was evaluated on a labeled dataset of oil sands video frames, achieving 90% accuracy compared to human labels.