Modelling calibration uncertainty in networks of environmental sensors

TL;DR

This paper introduces a variational method to model and quantify calibration uncertainty in environmental sensor networks, improving calibration transfer and enabling better deployment of low-cost sensors and citizen science.

Contribution

It presents a novel variational approach for calibration with uncertainty modeling, outperforming existing multi-hop calibration methods and extending to categorical data.

Findings

The method performs better than state-of-the-art calibration techniques.

It provides uncertainty quantification for sensor calibration.

Applicable to both continuous and categorical sensor data.

Abstract

Networks of low-cost sensors are becoming ubiquitous, but often suffer from poor accuracies and drift. Regular colocation with reference sensors allows recalibration but is complicated and expensive. Alternatively the calibration can be transferred using low-cost, mobile sensors. However inferring the calibration (with uncertainty) becomes difficult. We propose a variational approach to model the calibration across the network. We demonstrate the approach on 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 produced by citizen-scientist labelling. In Summary: The method achieves uncertainty-quantified calibration, which has been one of the barriers to low-cost sensor deployment and citizen-science research.