Cross-sensitivities: what confuses each sensor

No physical sensor reads only what its label says. Every instrument has cross-sensitivities, responses to inputs other than the target, and a useful dashboard knows about them and corrects for them. Here are the practical ones for Terrestream.

The SEN66's VOC index is moderately humidity-sensitive: at very high RH (above 80%), the index rises mildly even with no actual VOC change. The algorithm partially compensates, but on a sudden humidity step (say, a hot shower) the VOC reading will tick up. Terrestream does not adjust the VOC index for humidity, so read a VOC bump that arrives with a shower's humidity step with that in mind. Just as important is what the VOC index is not: it is non-specific and cannot isolate individual compounds. Terrestream does not measure, report, or detect formaldehyde or benzene, and must not be used to detect either (see formaldehyde and benzene).

The NO_x index responds to any oxidizing gas, primarily nitrogen oxides as intended, but also ozone (rarely high indoors but possible) and some halogenated compounds. Outdoor ozone infiltrating during a high-O₃ afternoon can produce a small NO_x-index rise that has nothing to do with combustion sources.

The PM optical sensor measures scattering. Anything that scatters infrared light at the wavelengths used will register: water droplets (fog, steam), some cooking aerosols (oil mist) above true PM. A steamy bathroom or boiling-pasta kitchen produces inflated PM readings that decay quickly as the droplets evaporate. Fonollosa et al. examine drift in metal-oxide gas sensor arrays and how to correct for it.

How Terrestream handles this: before it flags particles, it corrects PM_2.5 for humidity and holds back several particle alerts when high humidity alone would explain the rise, and when VOC and particles rise together in humid air, the notification adds that humidity may be affecting the particle reading. It also waits for a minimum NO_x index before pointing to a combustion source. Recommendations in the app carry a confidence level (high, medium, or low) that drops when outdoor data is stale or occupancy has to be inferred.