Efficient diagnosis of emissions from combustion processes plays a key role in their control, an essential part of the overall effort to mitigate the increasing greenhouse effect. In industrial furnaces, a set of sensors (CO x , SO x , NO x ) at the exhaust is used to monitor pollutant rates, thus providing the necessary information for control purposes. In the case of natural gas furnaces, measurements of O 2 and CO 2 contents are used to check the condition of the combustion process. In this work, we propose a method to estimate the O 2 and CO 2 contents at the exhaust of a natural gas prototype furnace from images of flames grabbed by a charge-coupled device (CCD) camera. Feature vectors obtained from computer processing of the grabbed images are used as input data to identify auto-regressive moving average (ARMAX) "black box" models having CO 2 content as output. Estimates of O 2 content by a Kalman filter running a preliminary ARMAX model help the overall performance of the method. Results show that the flame dynamics identified model is capable of yielding statistically significant estimates of both O 2 and CO 2 composition in the flue gas up to 10 s before the arrival of actual O 2 measurements. This outcome suggests that the inclusion of the proposed method in the closed-loop control strategy of similar combustion processes might be advantageous.