Modelling calibration uncertainty in networks of environmental sensors

Networks of low-cost sensors are becoming ubiquitous, but often suffer from low accuracies and drift. Regular colocation with reference sensors allows recalibration but is often complicated and expensive. Alternatively the calibration can be transferred using low-cost, mobile sensors, often at very low cost. However inferring appropriate estimates of the calibration functions (with uncertainty) for the network of sensors becomes difficult, especially as the network of visits by the mobile, low-cost sensors becomes large. We propose a variational approach to model the calibration across the network of sensors. We demonstrate the approach on both synthetic and real air pollution data, and find it can perform better than the state of the art (multi-hop calibration). We extend it to categorical data, combining classifications of insects by non-expert citizen scientists. Achieving uncertainty-quantified calibration has been one of the major barriers to low-cost sensor deployment and citizen-science research. We hope that the methods described will enable such projects.