A nationwide evaluation of crowd-sourced ambient temperature data

Growing concerns about heat in urban areas paired with the sparsity of weather stations have resulted in individuals drawing on data from citizen science sensor networks to fill in data gaps. In the past decade, a proliferation of crowd-sourced sensors has provided low-cost local air quality and temperature, with one brand having over 14,000 sensors deployed in the United States between 1 January 2017 and 20 July 2021. Although the air quality data from PurpleAir sensors have been widely studied, less attention has been paid to reported temperature. Gridded modeled temperature datasets are widely used in epidemiologic studies. The spatial granularity of the crowd-sourced sensor data captures local temperature variation which existing gridded datasets cannot, and can potentially be used to generate exposure assessments for health research. We compare temperature metrics reported by the dominantly used crowd-sourced sensor in the United States with a gridded temperature product, the North American Land Data Assimilation System (NLDAS)-2, which although not a gold-standard measure of temperature, is widely used in epidemiologic research. We evaluate the lag between indoor and outdoor sensor temperatures. We report associations of the difference between outdoor sensor temperatures and NLDAS-2 temperatures, an indicator of degradation, and the duration of sensor operation. Finally, based on the temperature range recorded by the outdoor sensors vis-a-vis NLDAS-2 temperatures, we provide a list of 271 (2.5%) sensors potentially misclassified as outdoor and likely located indoors. We observed that the outdoor sensors agreed well with NLDAS-2 (R 2 > 0.82). This association broke down under warm conditions (daily average NLDAS ≥21.1 o C). Our comparison suggests that a radiative-correction needs to be applied to use crowd-sourced data reliably. However, the spatial granularity of the continental sensor network can reduce the measurement error in exposure assignment compared to the NLDAS-2. Indoor sensor temperatures lagged hourly NLDAS temperatures by 2 hours across almost all climate zones. The mean difference in hourly sensor and NLDAS-2 temperatures increased by 0.57 o C for every operational year, suggesting that careful attention must be paid to degradation. Overall, we found that researchers should be aware of the limitations in crowd-sourced sensor air temperatures when examining extreme heat, or when aggregating sensor data across multiple years.