Monitoring mental workload in a fast and accurate manner is important in scenarios where the full attention of humans involved is fundamental to the security of others. Firefighters, air traffic controllers, and first responders are constantly submitted to such tasks. In many cases, in addition to a demanding mental task, humans are also under varying levels of physical strain. Measuring mental workload under such scenarios is challenging, especially when relying on wearable sensors. In this paper, we explore the combination of an automated artifact removal algorithm with spectro-temporal features for mental workload assessment "in-the-wild," where varying levels of physical strain are present. Experiments show these features outperforming classical spectral ones for mental workload classification under two activity types (biking and walking/running) and three activity levels (none, low, high). Improved performance was achieved when both feature types were combined, thus suggesting complementarity for the task at hand.