Machine learning techniques to improve the field performance of low-cost air quality sensors

Abstract

Abstract. Low-cost air quality sensors offer significant potential for enhancing urban air quality networks by providing higher-spatiotemporal-resolution data needed, for example, for evaluation of air quality interventions. However, these sensors present methodological and deployment challenges which have historically limited operational ability. These include variability in performance characteristics and sensitivity to environmental conditions. In this work, we investigate field “baselining” and interference correction using random forest regression methods for low-cost sensing of NO2, PM10 (particulate matter) and PM2.5. Model performance is explored using data obtained over a 7-month period by real-world field sensor deployment alongside reference method instrumentation. Workflows and processes developed are shown to be effective in normalising variable sensor baseline offsets and reducing uncertainty in sensor response arising from environmental interferences. We demonstrate improvements of between 37 % and 94 % in the mean absolute error term of fully corrected sensor datasets; this is equivalent to performance within ±2.6 ppb of the reference method for NO2, ±4.4 µg m−3 for PM10 and ±2.7 µg m−3 for PM2.5. Expanded-uncertainty estimates for PM10 and PM2.5 correction models are shown to meet performance criteria recommended by European air quality legislation, whilst that of the NO2 correction model was found to be narrowly (∼5 %) outside of its acceptance envelope. Expanded-uncertainty estimates for corrected sensor datasets not used in model training were 29 %, 21 % and 27 % for NO2, PM10 and PM2.5 respectively.