3.1. Gas Sensor Performance
CO calibration models for individual sensors exhibited high correlation coefficients (r
2 = 0.99), indicating strong statistical relationships between sensor readings and calibration gas concentrations. A global calibration model using data from all sensors resulted in a poorer fit (r
2 = 0.94,
Table 1) and greater absolute and relative error rates than individually calibrated sensors.
Span coefficients for CO ranged from 1.48 ± 0.01 SE to 1.52 ± 0.01 SE with a global model span coefficient of 1.50 ± 0.10 SE, where tolerances are denoting standard errors (SE). Offsets ranged from 7.63 ± 0.35 SE to 17.77 ± 0.15 SE ppm with a global offset of 12.12 ± 2.14 SE ppm. Absolute error across individually calibrated sensors ranged from 0.02 ± 0.38 SD ppm at 16 ppm to 0.13 ± 0.33 ppm at 3.2 ppm with an overall absolute error of 0.01 ± 0.45 SD ppm, where tolerances are denoting standard deviations (SD). Standard deviation ranged from 0.33 ppm at 3.2 ppm to 0.50 ppm at 32 ppm for all sensors calibrated individually (
Figure 3). Percent relative error across individually calibrated sensors ranged from 1.5 ± 0.4 SD% at 32 ppm to 9.7 ± 4.9 SD% at 3.2 ppm with an overall rate of 4.0 ± 4.0 SD% (
Figure 3). In all cases, the relative error rates for Spec Sensors DGS-CO 968-034 CO sensor performed better than the manufacturer stated measurement error rate of ±15%. The Monitor Labs Inc. CO Analyzer Model 8830 exhibited improved accuracy over our low-cost CO sensors under calibration, with relative error rates ranging from 1.0 ± 0.7 SD% at 32 ppm to 8.3 ± 5.8 SD% at 3.2 ppm with an overall error rate of 2.9 ± 3.9 SD%. The relative difference between the calibrated CO Analyzer Model 8830 and the DGS-CO 968-034 CO sensors was 3.3 ± 4.4 SD%. Both absolute and relative error rates for the DGS-CO 968-034 CO sensor were greater than those of the Citicell electrolytic CO sensor used by Ward et al. [
35,
36], but similar, if not slightly better, than those of the EC4-500-CO sensor. It is important to note that the range of concentrations in our calibration was smaller than those used in calibrating the EC4-500-CO [
16]. Using the global calibration model, the mean absolute error was approximately double and the SD was ~10 times greater than that found in individual sensor calibrations and ranged from 0.05 ± 2.81 SD ppm at 16 ppm to 0.26 ± 2.66 ppm at 24 ppm with an overall absolute error of 0.03 ± 2.80 SD ppm. The relative error was substantially higher using the global calibration model, ranging from 7.3 ± 3.9 SD% at 32 ppm to 80.0 ± 38.0% at 3.2 ppm with an overall absolute error of 27.9 ± 33.1 SD% (
Figure 3). Overall, the Spec DGS-CO 968-034 sensor performed well when calibrated individually, but variation between sensors resulted in high absolute and relative error rates, suggesting that the use of a global calibration model is inappropriate for this sensor.
CO
2 calibration models exhibited high correlation coefficients for both individual and global models (r
2 = 0.99,
Table 2), and equivalent absolute and relative error rates, suggesting that individual and global calibration models are appropriate for use with the Senseair K-30 CO
2 sensor. Span coefficients ranged from 0.695 ± 0.007 SE to 0.710 ± 0.011 SE with a global model span coefficient of 0.701 ± 0.005 SE. Offsets ranged from 143.947 ± 27.171 SE to 157.002 ± 18.807 SE ppm with a global offset of 148.439 ± 13.374 SE ppm. Absolute error across individually calibrated sensors ranged from 18.1 ± 5.5 SD ppm at 2000 ppm to 36.9 ± 3.7 ppm at 400 ppm with an overall absolute error of 0.01 ± 30.9 SD ppm. Standard deviation ranged from 1.9 ppm at 1000 ppm to 8.7 ppm at 3000 ppm for all sensors calibrated individually (
Figure 3). Absolute error rates using the global calibration model were nearly identical; SD was nominally greater, ranging from 18.1 ± 18.1 SD ppm at 2000 ppm to 36.8 ± 9.6 ppm at 400 ppm with an overall absolute error of 0.01 ± 36.7 SD ppm. Relative error across individually calibrated sensors ranged from 0.9 ± 0.3 SD% at 2000 ppm to 9.1 ± 0.9 SD% at 400 ppm with an overall rate of 2.8 ± 3.2 SD% (
Figure 3). Similarly, relative error when using the global calibration model ranged from 0.9 ± 0.9 SD% at 2000 ppm to 9.1 ± 2.4% at 400 ppm with an overall absolute error of 2.8 ± 3.4 SD% (
Figure 3). Overall concentrations, and absolute and relative percent error rates for the Senseair K-30 CO
2 sensor were within manufacturer stated measurement error rates; however, error rates at our lowest concentration (400 ppm) were greater than the manufacturer stated error rate (± 30 ppm or ± 3% of the measured value, whichever is greater). For comparison, the Li-COR LI-7000 exhibited favorable relative error rates ranging from 0.1 ± 0.1 SD% at 3000 ppm to 2.6 ± 0.4 SD% at 400 ppm with an overall error rate of 1.0 ± 0.9 SD%. The relative difference between the calibrated LI-7000 and K-30 CO
2 sensors was 3.7 ± 4.6 SD%. Absolute and relative error rates for the K-30 sensor were greater than those found for the DX6220 CO
2 sensor [
16] and the Valtronics (model 2015 BMC) CO
2 sensor employed by Ward et al. [
35,
36]. Similar to the findings by Yasuda et al. [
37], no temperature dependence of the offset was observed for the CO
2 sensor at ambient concentrations and for a temperature range of 10 to 40 °C.
Response time was evaluated as the time required for each sensor to shift from background to 90% of the applied concentration (T
90). CO sensor T
90 increased asymptotically with the concentration change and ranged from 32.3 ± 4.5 SD seconds with a concentration increase from 3.2 to 8 ppm and 107.0 ± 4.4 SD seconds with a concentration increase from 3.2 to 32 ppm (
Figure 4), substantially slower than the T
90 reported by the manufacturer and that reported for a EC4-500-CO sensor [
16] and a Citicell electrolytic CO sensor [
35,
36]. The CO
2 sensor T
90 response time also increased asymptotically and ranged from 41.0 ± 1.7 SD seconds when the concentration increased from 400 to 1000 ppm and 95.0 ± 2.6 SD seconds when the concentration increased from 400 to 4000 ppm (
Figure 4). T
90 for the K-30 CO
2 sensor was slower than the 20 s diffusion time reported by the manufacturer, but slightly faster than the response observed by Yasuda et al. [
37]. In comparison to other NDIR sensors, the response time of the K-30 was slower than the Valtronics (model 2015 BMC) sensor [
35,
36] and the DX6220 CO
2 sensor [
16]. A slow sensor response may complicate sensor use for the identification of peak gas concentrations and to fully describe the temporal variation in gas concentrations, but are not expected to hinder estimates of total CO and CO
2 fluxes in the smoke plume [
16].
3.2. Smoke Emissions Sampling
The performance of our smoke sampling module was demonstrated onboard an sUAS at Tall Timbers Research Station, FL. We collected five canister samples for VOC determination by remotely triggering the valve system at the 50–100 m range; two under prefire ambient conditions and three from the smoke plume during active burning. Regional burning outside of our study area may have contributed to trace VOC concentrations in samples. Prefire ambient samples contained total VOC concentrations of 0.8 and 1.8 ppbv, whereas samples collected during active burning contained 7.3 to 24.3 ppbv (
Figure 5). The seven most abundant VOCs observed in ambient samples were iso-pentane, benzene, 1-butene + isobutene, isoprene, α-pinene, n-decane, and n-octane. Samples collected from the smoke plume included a total of 42 VOCs. Six VOCs (i.e., iso-pentane, benzene, 1-butene + isobutene, 1,3-butadiene, toluene, and styrene) accounted for ~71% and 15 VOCs accounting for ~90% of total VOCs by ppbv observed during active burning (
Figure 5,
Supplementary Table S2).
Figure 5 presents the top 10 most abundant individual VOCs measured in the ambient air and in the fire smoke plumes. Emission factors were not determined because canister sample concentrations were too low for additional laboratory analysis of total carbon content.
Comparing the C4–C10 VOCs in the fire plumes with those reported in the literature, we found that the top 10 VOCs in our research were observed and reported in other studies [
13,
38,
39,
40,
41]. Dreessen et al. [
40] studied emissions transported from a Canadian wildfire that occurred in June 2015. Even though the plume was aged and diluted, high concentrations of isoprene, benzene, and xylenes were observed. Similar to our results, α-pinene, benzene, toluene, 1-butene, isoprene, 1,3-butadiene, and m+p-xylenes were in the top 10 VOCs (>C4) in a fresh Canadian boreal forest fire plume [
41]. Simpson et al. [
41] reported 0.99 ± 032 ppbv of benzene and 0.37 ± 0.14 ppbv of toluene; these values are a factor of 1.2 to 3.5 lower than our levels (
Figure 5). This is most likely due to different meteorological conditions and variation in the dilution of fire emissions during the sampling in fire plumes. In general, the VOC results presented here show good agreement with the literature.
3.3. Smoke Emissions Sensing
The performance of our smoke sensing module was demonstrated in a near-ground (8-m height above ground) deployment of the instrument at Sycan Marsh Preserve, OR, USA. Prior to the burn, CO and CO
2 sensors were calibrated in the laboratory over a concentration range spanning 3 to 48 ppm for CO and 400 to 6000 ppm for CO
2. Calibrated CO concentrations in the smoke plume ranged from an ambient minimum of −1 to a peak of 197 ppm (
Figure 6); boxcar smoothing constrained the CO concentration range from −1 to a maximum of 180 ppm. Calibrated CO
2 concentrations ranged from an ambient minimum of 440 ppm to a maximum of 6326 ppm (
Figure 6) and boxcar smoothing constrained the CO
2 concentration range from 418 ppm to a peak of 4336 ppm. The ranges of the CO
2 and CO concentrations observed at Sycan Marsh Preserve indicate that future deployments will require calibration of the CO sensor over a greater concentration range while the CO
2 sensor’s calibration range was appropriate for the application. MCE estimates the efficiency of fuel combustion with smoldering combustion recognized as occurring below a threshold of 0.85 to 0.90 and flaming combustion occurring above this level. Computed using the boxcar smoothed data, the MCE measured in Sycan Marsh Preserve, OR ranged from 0.84 to 1.00 (
Figure 6) and generally matches findings from studies that observed flaming combustion in senescent grass and pine litter fuel types [
8,
13].
PM
2.5 mass concentration measurements using the Plantower PMS5003 sensor ranged from 0.0 to 1717 µg m
−3 and from 0.8 to 1599 µg m
−3 after boxcar smoothing (
Figure 6). As expected, levels observed directly in the smoke plume here were greater than other studies that used the sensor to investigate regional smoke impacts on ambient air conditions during the fire season [
24,
26]. Recent evaluations of the Plantower PMS series (i.e., PMS1003/3003/5003/7003) sensors primarily investigated the sensor at PM
2.5 concentrations up to 150 µg m
-3 and found strong correlation with reference monitors [
23,
24,
25], but one study found correlations were degraded during the wildfire season [
24]. Multiple authors noted that raw sensor readings may overestimate PM mass concentrations [
23,
24,
25] and attributed some overestimation to an increased sampling frequency over reference monitors that permitted greater detection of temporal variation and short lived events [
25,
26]. Authors also observed a nonlinear response at high concentrations (pronounced above 125 µg m
−3, [
23]) and a high correlation with relative humidity [
23,
25]. Percent error after calibration ranged from 21% to 201% in one study and varied depending on the type of reference monitor used for calibration [
23]. PM
2.5 mass concentrations observed in our study were very high in comparison with other studies and exceeded the manufacturer’s suggested maximum concentration (i.e., 1000 µg m
−3). Therefore, we urge caution in interpreting PM
2.5 mass concentrations observed directly in smoke plumes using uncalibrated Plantower PMS5003 sensors but highlight the potential utility of investigating temporal trends using the sensor. Further exploration of sensor calibration and environmental correction are necessary to ascertain the sensor’s measurement limits in smoke plumes.
Air temperature and relative humidity are essential parameters for predicting fuel flammability and fire behavior, and are routinely collected and used in fire behavior, fire monitoring, and fire modeling studies. Data collected during the experimental burn illustrate the dynamic range of environmental conditions observed with the passage of a flaming front. At 8 m height, the temperature ranged from 11.5 to 50.5 °C and the relative humidity ranged from 8.6% to 43.6% (
Figure 7). Further work is necessary to compare the temperature and humidity sensors used in our sensing instrument with industry standards in fire weather forecasting (i.e., Remote Automated Weather Stations (RAWS)) and to better understand how gas and particulate sensors perform across such a large range of environmental conditions at short time scales.