Pixels to pyrometrics: UAS-derived infrared imagery to evaluate and monitor prescribed fire behaviour and effects

UAS platform and imagery collection

Prior to ignition, we launched the UAS to collect images of the fire behaviour. We positioned the UAS directly overhead the centre of the plots (described in section 'Prescribed fire imagery aquisition and analysis'), facing north and at 122 m above ground, and collected thermal imagery at 0.2 Hz (5 s interval) to capture ignitions and post-frontal fuel consumption. All UAS imagery was collected with an Autel Evo II 640t UAS, featuring a 48-megapixel colour camera and a 0.32-megapixel IR camera. The thermal camera is an Infiniray Micro III Lite 640 (Yantai, Shandong, China) uncooled radiometric sensor that samples at the longwave IR wavelengths, 8–14 μm with a temperature range of 0–650°C. At 122 m, the colour and infrared sensors ground sampling distance are 1.2 and 14 cm pixel⁻¹, respectively. With a realistic flight time of 30 min, five batteries, and field charging array, we could perpetually hover the UAS over the fire. About every 30 min, we paused thermal image sampling to swap the UAS battery, returning to the same altitude and location to resume monitoring. It is important to note we used the low gain setting to capture infrared images, which is less sensitive than the high gain but can sense at the full operational range of the camera. Saturation refers to the highest temperature at which a sensor can accurately measure. Above this point, the sensor cannot differentiate between temperatures and may provide inaccurate readings. Sensor saturation can lead to underestimation of the energy produced, which is critical to precisely estimating energy release of combusting biomass via IR sensing. The IR sensor used in this project is certified by the manufacturer to 650°C, below the maximum potential combustion temperature (Paugam et al. 2013; Moran et al. 2019) and is subject to the dependency of flame emissivity on flame depth (Johnston et al. 2014), but we selected this camera because of its affordability and simple use. While the camera limitation should be acknowledged, most prescribed fire typically involves low intensity fire, which is well below the maximum combustion temperature and IR saturation temperature (Iverson et al. 2004; Kennard et al. 2005).

Thermal imagery preprocessing

The thermal imagery acquired from the UAS in raw form was a 16-bit 1-band raster, with no geolocation information, requiring appropriate processing for quantitative spatial analysis. The three steps used to process the imagery are: (1) conversion to temperature (see Supplementary material Eqn S1); (2) stabilisation; and (3) georectification. All pyrometrics were derived from IR imagery preprocessed using the methods described, similar to Moran et al. (2019, 2022) and using equations developed and applied by O’Brien et al. (2015) and Hudak et al. (2016). After testing four stabilisation algorithms and the use of a mask to remove burning pixels, we determined an Enhanced Correlation Coefficient algorithm without a fire-mask was the best stabilising model for this study, used for all imagery hereafter (Fig. S2). The end result of the pre-processing were georeferenced, temperature-calibrated image stacks, where each pixel represents fire radiative power (FRP, W m⁻²) at a specified location. Each image stack can be thought of as a time-lapse of fire behaviour. FRP is then converted to fire radiative energy (FRE, MJ m²) by taking the integral of FRP over time. For more details on thermal imagery preprocessing, see Supplementary material S1.

Experimental calibration burn

We conducted an experimental burn to validate the estimates of surface fuel consumption derived from a UAS-mounted IR sensor. To do this, we burned known quantities of ponderosa fuel, measured the consumption directly by weighing fuel before and after the burn (hereafter weight-based), and compared it to consumption estimates derived from the UAS-IR imagery (hereafter UAS-based). Images were processed as described above to generate FRE of each plot. We then calculated fuel consumption using two published estimators (Eqns 1 and 2) and compared it with consumption calculated by weighing fuel before and after. This calibration burn compared the UAS-IR method of estimating consumption with a validated, ground-truth counterpart in a variety of fuel conditions.

We precisely weighed and homogeneously arranged variable amounts of dry ponderosa litter and cut logs (<15 cm in diameter) in 10 1-m² plots. We selected coarse woody debris (CWD) and litter quantities that are representative of fuel conditions encountered in ponderosa forests, (Table 1). Roccaforte et al. (2012) reported a maximum 10 and 0.8 kg m⁻² of CWD and litter loadings, given the bulk density of litter is ~4 Mg ha⁻¹ cm⁻¹ (Stephens et al. 2004). We doubled each of these quantities to account for the spatial heterogeneity typical of surface fuels. The design of the 10 plots was to burn across a range of fuel loads at different ratios of litter and wood. We first stratified our plots by wood weight (including moisture-related weight), arranging wood from 2 to 20 kg at 2-kg intervals at each plot. To ensure we sampled variable litter and wood ratios that were representative of the sites, we then added ponderosa needle cast collected nearby the calibration plot at weights of 0.5 kg, 1 kg, and 2 kg (Table 1). Fuel moisture was sampled immediately prior to ignitions from three wood and three litter samples and calculated as the relative loss of water content after drying in an oven at 70°C for 24-h intervals until the dry weight was stable (Keane 2015).

Plots were ignited with a flare to avoid adding fuel and energy to the plot during ignition. We collected IR imagery at 0.2 Hz, 61 m above ground for 2 h and 15 min. Consumption was immediately stopped by spraying all the plots with water and extinguishing any smouldering. After the burn, all remaining ash and charred wood was collected, dried until the weight stabilised, and weighed to calculate the total consumption following eqn 3 in Smith et al. (2005). This approach does not account for incombustible material in the fuel, the proportional combustion completeness, and potential loss of ash from convective heat and wind. A solution is to use equation 12 from Smith et al. (2005) that accounts for the entire combustion budget by calculating the loss on ignition of pre- and post-fire fuel components. Nonetheless, the simple subtraction method used here calculates fuel consumption to ~10% accuracy of the more time-consuming approach (Smith et al. 2005).

We compared methods for fuel consumption (FC) estimation using two different FRE-based approaches (Smith et al. 2013; Hudak et al. 2016):

where FC is consumption (kg m⁻²), hc is heat content (MJ kg⁻¹), rf is radiative fraction, and W_c is water content (%). Eqn 1 is an energy density-based approach that accounts for the heat content of the fuel and the proportion of energy released radiatively (Hudak et al. 2016). We used an rf of 17%, the mean radiative fraction of the 12 Quercus litter and woody debris experimental burns carried out by Kremens et al. (2012). The average heat content for all ponderosa fuel components was 20.86 MJ kg⁻¹ (Van Wagtendonk et al. 1998). Equation 2 comes from Smith et al. (2013). This moisture-based approach does not account for the effect of variable heat contents of fuel types and instead addresses the strong effect moisture has on radiative fraction and consumption (Smith et al. 2013).

Prescribed fire imagery acquisition and analysis

We collected data at three operational prescribed fires: two in Arizona; and one in Florida (Table 2). The prescribed fire sites were all composed of a fire tolerant Pinus species but displayed substantial variability in fuel loading, fire seasonality, and target fuel load reductions (Table 2). To collect data on prescribed fires in Arizona, we obtained permission from the USDA Regional Fire and Aviation Office to operate UAS and coordinated with local US Forest Service districts to attend approved prescribed fires. In Florida, the prescribed fire was on private land, so we coordinated with landowners to attend the fire. We adhered to the Federal Aviation Administration Part 107 rules and the safety regulations written in the US Forest Service research contract (when attending agency-administered prescribed fires). Sites were selected based on availability of prescribed fires during the 2022–2023 season. Prescribed fires were ignited in the Fall, Winter, and Spring and were within the target weather window described in the prescribed burn plans (see Supplementary material).

Prior to ignitions, we located plots that were representative of the fuel conditions of the prescribed fire sites. We placed plots near the perimeter of the fire, typically on opposite sides of the burn unit to capture multiple plots during one ignition operation. Plot size of the thermal imagery was variable, dependent on the UAS altitude, drift, and stability. Generally, plots were about 0.4 ha (Table 2). We placed four 40-cm diameter aluminium pans in each plot, one in the centre and three in the north, south, and west directions about 15–20 m from the centre pan (prioritising canopy openings to avoid occlusion). The pans acted as ground control points (GCPs), allowing us to continually fly the UAS over the same location, georectify images, and validate the georectification. Shiny aluminium contrasted the forest floor well and had a low emissivity, meaning the plates were clearly visible in the infrared and colour imagery. We used the pans to georectify the thermal image stacks and verify that the stabilisation and coregistration workflow produced acceptable results, as described below. At the Hanna Hammock Rx, we geolocated each pan using a real-time kinematics GPS with centimeter-level accuracy. At the Hundred Rx and Wild Bill Rx, we measured the distance between each GCP and inferred their coordinate position by orthorectifying pre-burn images of the plot via photogrammetry (Agisoft 2023). This method was centimetre-level accurate as well, but the geolocation was subject to the precision of the UAS GPS. We recorded the average daily fuel moisture of 10-h fuels from the nearest Remote Automated Weather Station, available from the Western Regional Climate Center (http://www.raws.dri.edu).

We used the best performing method from the calibration burn and the moisture-based approach (Eqn 2), to convert FRE to consumption for the prescribed fire plots. One challenge with monitoring with UAS is capturing post-frontal combustion (Moran et al. 2022), which may continue for hours or days after the burn. We balanced collecting as many plots as possible with capturing all possible post-frontal consumption by sampling about two plots per operational shift, each plot being sampled until radiant energy release had substantially subsided. To estimate post-frontal combustion, we fit a linear regression with a logarithmic transformation of the plot FRP at each infrared image_i in the image stack. Linear regressions were calculated for each burn:

where t is time. We fit the linear regression to the energy release from the peak energy release to the end of the image capture, which visually exhibited a logarithmic relationship. We then used this model to predict unmeasured energy release of the plots from when data collection stopped to when the theoretical energy release ≤1 W m⁻². Comparing the total energy (FRE) measured with the predicted unmeasured energy release gave us a corrective factor, which we applied to each consumption raster to account for any unmeasured energy. For example, if the regression predicted 20% of energy was not captured during collection, then each pixel would be multiplied by 1.2.

To estimate rate of spread, we used the masked FRP raster stack to generate an arrival time, or progression, raster of the fire when each pixel reached flaming power, 1070 W m⁻² (Hudak et al. 2016). The rate of spread was then derived as the rate of change or slope (radians) of the progression raster (Moore et al. 2020) using the equation:

where ROS is measured in m min⁻¹. The slope raster was then converted to points and interpolated for all progression pixels using inverse distance weighing. Residence time was calculated as the total time (in seconds) that each pixel was greater than or equal to 300°C (Wotton et al. 2012).

Statistical analysis

To evaluate the effectiveness of UAS-IR imagery at capturing the ‘signal’ of prescribed fires, we assessed whether the spatial resolution could adequately characterise fire behaviour variability. In signal processing, the Nyquist criterion states the sampling interval (or resolution) should be twice that of the frequency in the sample (Shannon 1949; Hengl 2006). Therefore, we evaluated the resolution of the ‘signal’ to determine the resolution needed of the sensor. We constructed a semi-variogram of each plot’s FRE raster to determine the spatial scale where autocorrelation was no longer significant. Then, we calculated the median semi-variogram range of all the plots and resampled each pyrometric raster to this value. We confirmed that autocorrelation was reduced with a Moran's I-test. These spatially-independent samples of all three pyrometrics were then compared to assess differences among prescribed burns with a non-parametric Kruskal–Wallis test. If the Kruskal–Wallis test indicated significant differences, we used the Dunn test, a post hoc test for non-parametric data, to determine specifically which burns were different from one another.

UAS-derived consumption

Observed weight-based consumption of the 10 1-m² plots ranged from 2.1 to 19 kg m⁻² with all plots consuming 95% or more of dry weight fuel (Table 1). We measured average litter and wood fuel moisture as 11%. UAS-based consumption estimates were in high agreement with observations across all plots (Fig. 1). Converting radiative energy using either Eqns 1 or 2 to consumption was highly correlated with observed consumption. However, using equation 1 from Hudak et al. (2016) consistently underpredicted consumption. The linear regression of observation as the response had a slope of 0.77 (R² = 0.99). Converting FRE using the fuel moisture approach from Smith et al. (2013), Eqn 2 was closer to a 1:1 relationship with a linear regression slope of 1.1 (R² = 0.99). Therefore, all estimates of consumption for the prescribed fires (Table 2) were generated with the Smith et al. (2013) equation (Eqn 2).

Fig. 1.

Calibration burn results from weight-based measurements of consumption and UAS-based consumption (left plot). Brown points (triangles) use Eqn 1 from Hudak et al. (2016) and green points use Eqn 2 from Smith et al. (2013). Linear regressions of observed as a function of predicted (observed ~ predicted consumption, solid lines) reveal a slight overprediction when applying Eqn 2 and an underprediction when applying Eqn 1. The density plot (right plot) highlights the variability of fire radiative energy (FRE) pixels of plot 1 (left image) and 10 (right image), with remaining plots represented as lines.

UAS-IR fire behaviour

Extracted pyrometrics from the UAS-derived IR imagery agreed with qualitative ocular observations of the burns: Hanna Hammock spread the fastest; and Hundred and Wild Bill prescribed fires consumed more fuel with longer residence times (Fig. 2). Predicted FRE using Eqn 3 to account for unmeasured FRP performed well, with strong model fits with the captured energy (Fig. 3). Plots where the UAS imagery collection captured enough post-frontal combustion that a decay curve was evident had the strongest model performance. At the Hanna Hammock plots, majority of the energy was released within 10-min, resulting in most energy being directly measured and only a minor fraction modelled with Eqn 3. The prescribed burns in Arizona continued to release significant energy for hours after the peak FRP and therefore the regression-based estimates of energy release were critical for estimating total energy released and consumption. Average total consumption for Hanna Hammock, Hundred, and Wild Bill were 0.3, 1.4, and 4.9 kg m⁻², respectively. The consumption data was heavily positively skewed (Fig. 4). Maximum within-plot consumption was 65 kg m⁻², sampled at plot 1 of the Wild Bill Rx. Median rate of spread was 2.5, 0.7, and 0.7 m min⁻¹ for Hanna Hammock, Hundred, and Wild Bill prescribed fires, respectively. Rate of spread reported by the UAS imagery ranged from 0.002 to 4.0 m min⁻¹, the maximum recorded at Plot 1 of Hanna Hammock and the minimum recorded in plot 4 of Wild Bill prescribed fire. Median residence time was 40, 96, and 170 s for Hanna Hammock, Hundred, and Wild Bill prescribed fires, respectively.

Fig. 2.

Location and images of fire behaviour of the prescribed fires monitored (a), selected IR images and derived rate of spread raster of the Hanna Hammock plot 1 (b), and selected IR images and derived consumption raster of the Wild Bill plot 3 (c). Hanna Hammock Rx had the highest visually observed rate of spread and lowest consumption. Wild Bill Rx had a large amount of coarse woody debris present in plots and had the highest visually observed estimate of consumption and residence time.

Fig. 3.

Fire radiative power (W m⁻²) over time of each of the burn plots. Red dashed line is the logarithmic transformed linear regression and respective R² values denote model performance evaluated with measured energy data. Dark shaded area is the energy measured by the UAS-IR imagery, lightly shaded area is unmeasured energy predicted using the regression. Measured and predicted energy values were summed in the final consumption estimates. The green fine-dashed line in Wild Bill 4 plot represents a long pause (~1 h) in imagery collection during minimal, static fire activity. This particular plot was burned in early morning when conditions were relatively cool and wet, so fire intensity was particularly low. Ignition operations then made a second pass at 150 min, corresponding to the spike in fire radiative power.

Fig. 4.

Violin distributions and boxplots of UAS-derived pyrometrics from each prescribed fire: Hanna Hammock Rx, Hundred Rx; and Wild Bill Rx. A significant Kruskal–Wallis rank sum test that two or more of the distributions are different is marked as an asterisk on the x-axis label. Post hoc Dunn tests to determine pairs of significantly different distributions are marked as asterisks on the y-axis label.

Using a semi-variogram to identify autocorrelation for each plot’s radiative energy (FRE) raster, we found the median range to be 4.6 m. Before resampling, FRE rasters of each plot had a median resolution of 0.13 m (Table 2) and Moran’s I-value of 0.96, indicating a very strong positive autocorrelation. 35 pixels were resampled to 4.6 m resolution to minimise autocorrelation. Therefore, the Nyquist sampling theorem would suggest a sampling resolution of 2.3 m or less is needed to characterise the fire behaviour of prescribed fires. After spatial resampling (to 4.6 m), median Moran’s I-value was reduced to 0.29: while spatial autocorrelation was still present, it was markedly reduced. A Kruskal–Wallis test of the resampled pyrometric data showed that each pyrometric had significantly different groups (Fig. 4). The post hoc Dunn tests showed that Wild Bill and Hundred groups were similar in terms of rate of spread or residence time distributions (Fig. 4), though all other Dunn tests were significant (P < 0.025).