In order to determine an acceptable period of sensor calibration that is affordable for network operators, but minimizes sensor inaccuracies, the onset of sensor drift and the appearance of spiking in temperature data needs to be quantified. For this purpose, a clean dataset of 3 years after sensor replacement was obtained from a South Alabama Mesonet station, located 2.1 km from Mobile Bay, Alabama, after rigorous manual review of daily timeseries plots. Data from the 4 temperature sensors were compared and statistics of sensor temperature differences were examined. In many cases, sensor drift starts to manifest itself after 2 years. As the sensors age, errors like spikes and dips also become more frequent. This seems to be especially prominent with the 9.5 m model 107 thermistor, possibly because of the longer cable required to connect the sensor to the datalogger.
During the first 2 years, the temperature differences between the 1.5 and 2 m sensor pair hover around zero and lie primarily between the sum of the manufacturer specified sensor accuracies of the two different sensors (i.e. ±0.7⁰C), with only outliers above and below those values. The nature of these outliers will be investigated and reported and could include nocturnal temperature inversions or transient larger temperature differences caused by daytime turbulent eddies. Over time, the spread in the data of these two sensor pairs increases and becomes primarily positive, i.e. the 1.5 m temperature sensor reads larger values even at night time.
For the 9.5 and 10 m sensor pair this trend is even more pronounced. Positive values (i.e. larger 9.5 m temperature measurements) are seen all 3 years, but are confined to the ±0.7⁰C interval during the first 2 years, except for some outliers. After that, more than 50% of the temperature differences are positive and the positive outliers become larger.
Aside from quantifying sensor drift, another important use of the above-mentioned temperature difference statistics is to define like-sensor test thresholds in an automated quality control (QC) scheme to be implemented in the future. The challenge is setting accurate QC flagging thresholds while preserving realistic phenomena like low-level temperature inversions, which are important to agricultural interests and have an impact on air quality. Low-level temperature inversions are very common in the north-central Gulf Coast area and cause large temperature differences to occur between all 4 temperature sensors during the night. For this reason, sensor comparisons were stratified by day and night.
While night-time inversions are common, they sometimes occur during the day as well. There are a few negative daytime outliers in the 2-10m temperature difference which could be caused by turbulent eddies during daytime mixing of the air and temporarily causing a negative temperature difference. Sometimes the passage of a cold or gust front shows a lag in one of the sensors leading to a momentarily large temperature difference.
Results from this study will be used to evaluate whether bias adjustments need to be made so that low-level temperature inversion climatologies can be established. Similarly, land-sea temperature differences used to study sea-breeze evolution may need to be bias-adjusted. Future work will include exploring the use of Artificial Intelligence techniques to address these issues of bias correction. Lastly, the results will be used to establish a quality-control scheme with the goal to identify sensor drift, but to not flag realistic temperature inversions.
This work was supported in part by NSF under Grant OIA-2019511
The data used in this research was supported in part by the National Weather Service's National Mesonet Program, which helps make these and other data available to forecasters, researchers, and the general public.

