This paper investigates whether mental workload can be classified in an operator setting using unobtrusive psychophysiological measures. Having reliable predictions of workload using unobtrusive sensors can be useful for adaptive instructional systems, as knowledge of a trainee’s workload can then be used to provide appropriate training level (not too hard, not too easy). Previous work has investigated automatic mental workload prediction using biophysical measures and machine learning, however less attention has been given to the level of physical obtrusiveness of the used measures. We therefore explore the use of color-, and infrared-spectrum cameras for remote photoplethysmography (rPPG) as physically unobtrusive measures. Sixteen expert train traffic operators participated in a railway human-in-the-loop simulator. We used two machine learning models (AdaBoost and Random Forests) to predict low-, medium- and high-mental workload levels based on heart rate features in a leave-one-out cross-validated design. Results show above chance classification for low- and high-mental workload states. Based on infrared-spectrum rPPG derived features, the AdaBoost machine learning model yielded the highest classification performance.