1. Introduction
The risk of large and severe wildfires has steadily increased throughout the western U.S., posing serious threats to human infrastructure and ecosystem values [
1,
2,
3,
4]. While much attention has focused on wildfire in forests, grass- and shrub-dominated ecosystems have shown capable of rapid fire spread and extreme fire behavior [
5,
6]. Although wildfire is a regular disturbance agent in semi-arid grasslands, the spread of non-native invasive grasses in the western U.S., such as lovegrasses (
Eragrostis spp.), buffelgrass (
Pennisetum ciliare), and brome grasses (
Bromus spp.) [
7,
8,
9], have increased fine-fuel loads that intensify wildfire activity [
10]. Wildfire, in turn, further promotes the spread of these invasive species, creating a positive feedback loop with wildfire [
11,
12,
13,
14] and elevating the risk to natural resource values compared to historic conditions [
15,
16,
17]. Woody plant encroachment has further changed wildfire regimes in western U.S. grasslands [
18,
19]. Shifts in the abundance of invasive grasses and woody plants, and high interannual variability in precipitation, influence fine-fuel production in ways that require up-to-date assessments of fuel loads across heterogeneous landscapes of the western U.S.
Changes in fuel and wildfire regimes present novel risks to highly valued assets including infrastructure, hydrologic function, critical habitat, and endangered or threatened flora and fauna [
20]. Altered and novel fire regimes, along with annual changes to the abundance and distribution of fine-fuels, present challenges to land managers implementing competing management objectives. Comprehensive fire plans that include prescribed fires and other fuel reduction techniques represent an accessible strategy for altering fuel loads, mitigating fire risk, and generally influencing ecological trajectories. With dramatic increases in wildfire activity throughout the western U.S., prescribed burning is a widely advocated for but under-used approach to fuels reduction [
21], with the potential to support resource priorities of land management agencies and concerns surrounding wildfire management [
22,
23]. Furthermore, some management units have records of prescribed and managed wildfires dating back decades [
23,
24,
25], yet the effect of these fuel treatments, and those of naturally occurring wildfires, on fuel load and their subsequent recovery remain largely unknown and undescribed in non-forested areas.
The Buenos Aires National Wildlife Refuge (BANWR;
Figure 1, [
26]) located in southern Arizona in the southwestern U.S., typifies a managed area that has experienced increased risk of wildfire from invasion of non-native grasses. Composed largely of semi-arid grasslands, BANWR has a storied history of land use and management, including the refuge’s primary mission to reintroduce the federally endangered masked bobwhite quail (
Colinus virginianus ridgwayi) into its historical range [
27]. In brief, commencing at the turn of the 20th century, there were periods of intensive livestock grazing resulting in near fire exclusion and shrub encroachment, followed by a cessation of grazing, and active fire management beginning with refuge establishment in 1985 [
26,
28]. Previous studies have addressed the growing concern of wildfire on the remaining masked bobwhite quail habitat and direct and indirect management actions to improve population viability of the reintroduced birds through fire management [
25]. The BANWR maintains an active prescribed burn program to ameliorate the threat of extreme wildfire events. The refuge is composed of over 80 habitat management units, with 60 units designated for inclusion in the prescribed burn plan that has operated since refuge establishment [
25]. Wildfires ignited by lighting associated with the North American Monsoon and anthropogenic sources play an active role in shaping ecosystem function in BANWR and the surrounding Altar Valley. Spatially explicit and updateable information on fine-fuels can assist wildlife biologists and fire managers in developing wildfire mitigation actions that weigh the trade-offs between quail conservation planning and fuel reduction.
Advances in remote sensing applications can capture interannual changes in fine-fuels at high-resolution in broadly distributed semi-arid grasslands [
22,
29]. The increased variety and availability of remotely sensed data provide a promising set of tools for assessing how vertical and horizontal fuel bed and canopy structure are altered by fire management actions [
30]. A comprehensive review of relevant literature from 1986 to 2019 [
22] of fuel and fire risk in the southwestern U.S. shows an increasing body of work related to remote sensing of invasive grasses and fine-fuels. Remote sensing assessment of fine-fuels can reduce resources required to monitor fuel loads in the field [
31] and is readily updateable to assess rapid changes in fuel condition. Remote sensing imagery that is scalable to detect the heterogeneity in fine-fuel loads in semi-arid grasslands may be particularly useful for assessing the effects of management actions and provide spatial inputs required by wildfire behavior models. We modeled fine-fuels over time (2015 to 2020) using the Sentinel-2A system, which produces imagery at multiple spatial resolutions (60 m, 20 m, and 10 m) and has an enhanced spectral resolution over earlier Landsat, Moderate Resolution Imaging Spectroradiometer (MODIS), and Satellite Pour l’Observation de la Terre (SPOT) sensors [
32] appropriate for detection of changes in fine-fuels. The advanced Sentinel-2A system provided an effective and rapid way for detecting, spatial modeling, and assessing changes of fine-fuels in a semi-arid grassland and in response to prescribed management treatments and wildfire at BANWR [
33].
Our objectives were to: (1) develop and validate a yearly model of fine-fuel, or herbaceous vegetation biomass, over a large grassland area and (2) assess changes to, and recovery times for, fine-fuels after wildfire and prescribed fire events. A quantitative understanding of fine-fuel return intervals provides fire managers with a means for understanding near- and long-term impacts of management actions and insight on treatment effects. In meeting our objectives, we demonstrate application of remote sensing to modeling fuel conditions in grasslands that are prone to future non-native plant invasion and climate change.
3. Results
Since 1985, prescribed fires accounted for 65% of fire events recorded at BANWR in the FMIS database (
n = 133, average size = 646 ha) and wildfires (
n = 71, mean size = 319 ha) accounted for the remaining 35%. Since 2015, 8 of the 12 fires recorded in the FMIS database were wildfires and 4 were prescribed fires (
Table A1). Of the eight wildfires, half were ignited by natural and half by anthropogenic sources. Three of the four prescribed fires were conducted in March and one was implemented in June. One of the eight wildfires occurred in May, one in June, two in July, three in August, and one in December (
Table A1). Our 2015 to 2020 study occurred during a period of fluctuating annual precipitation with record low annual precipitation in 2020 (
Figure 2a), high mean annual temperature (
Figure 2b), and high maximum annual vapor pressure deficit (
Figure 2c) in both the southern and northern part of the refuge, relative to the 1985 to 2014 historical record. Notably, in 2019, the study area experienced above average precipitation and cool temperatures that promoted low-fire behavior. Two wildfires were reported in 2019, both of which represented the smaller acreages burned during the study timeframe. Conversely, in 2020, the study area experienced record-setting low summer monsoon precipitation, above average temperatures, and high vapor-pressure deficit (
Figure 2).
Results from training of the 2015 fine-fuel model indicated an optimal RMSE of 500 kg/ha and an adjusted-R
= 0.41 for the final random forest model based on internal cross validation. The top five variables included in the 2015 training model included: band 11 (shortwave infrared) during vegetation dormancy, and band 5 (red-edge), band 4 (red), band 8 (near infrared), and band 6 (red-edge) during peak greenness of vegetation, as shown in
Table 1. Variable importance values indicated the percent increase in RMSE with predictors iterated in and out of the model. Of the 20 spectral variables included in the model, 12 of 20 were from the time period of peak greenness of vegetation while the other 8 were from vegetation dormancy. Total variable importance from spectral bands measured during peak greenness of vegetation accounted for 0.68 of overall importance, while spectral responses measured during vegetation dormancy accounted for 0.32. Subsequent spatial predictions for fine-fuel models from 2016 to 2015 were based on these variables.
When compared with the 20% reserved testing data that we withheld to validate and assess model accuracy (separate from internal model training cross validation), the 2015 fine-fuel model (
Figure 3a) had an adjusted-R
of 0.52 (F = 93.62,
p < 0.0001). The accuracy of the 2020 fine-fuel model based on its relation to the 2020 field-based measurements (R
= 0.63,
Figure 3b) was similar to that of 2015 and demonstrated the robustness of the model over time. The fine-fuel models provided continuous estimates of fine-fuels from lows of 200–400 kg/ha to highs of 2600–2800 kg/ha throughout BANWR (
Figure 4). The annual series of Sentinel-2A derived models was used to map the yearly abundance and distribution of fine-fuels (kg/ha) at a 10-m resolution for BANWR (
Figure 4 and
Figure 5). For all years, fine-fuel data layers generally showed increasing values in a north to south direction, which was associated with increased elevation, increased precipitation, and cooler temperatures in the southern portion of the study refuge (
Figure 2). The minimum and maximum amount of fine-fuel per year were generally within ±200 kg/ha of each other. Across the study area, fine-fuel was visibly reduced within fire perimeters following a fire event or treatment.
Close examination of two fire events, the 3-Hills wildfire (
Figure 5a–c) and 2016 Airport prescribed fire (
Figure 5d–f), showed high fine-fuel loads prior to ignition (
Figure 5a,d), with visible reductions in the first-year post fire (
Figure 5b,e). The 2016 Airport prescribed fire resulted in an average reduction of −660 kg/ha the first-year post fire, while the 3-Hills wildfire produced an average reduction of −480 kg/ha. The prescribed fire showed a homogeneous burn pattern and nearly complete overall reduction in fuels. Careful examination also revealed intentional and strategic prescribed burn patterning with greater reduction of fuels along ridge tops and benches, whilst lower reductions were achieved in areas of lower topography. The third-year post fire clearly showed a recovery of fuels, although not to the extent of pre-fire conditions (
Figure 5c,f).
When 2015 to 2020 fires were compared to controls by year, wildfires tended to show greater impacts to fine-fuels than prescribed fires with greater difference between averages (
Table A1) and greater differences between standard deviations (
Table A2) in estimates of fine-fuel. Annualized since time of fire, there were greater average reductions due to wildfire (−516 kg/ha) compared to prescribed fire (−152 kg/ha) in the first-year post fire (
Figure 6,
Table A3 and
Table A4). In contrast, unburned control plots in the first-year post fire showed virtually no change in fine-fuel estimates (−8 kg/ha). From 2015 to 2020, average values of fine-fuels estimated within burned areas tended to be lower (
Table A1), and with lower variation (
Table A2), than unburned reference plots. Notable exceptions to the expected reductions of fine-fuel included both the 2019 High Peak wildfire and the 2019 East Gate Border Patrol prescribed fire in which both had increases in estimates of fine-fuels the first-year post fire (
Table A4 and
Figure 6). Higher than expected fuel in 2019 was likely due to above average annual precipitation (
Figure 2), as even the unburned control plots increased by 220 kg/ha in 2019.
When grouped by years, the combined metrics of fuel reduction for prescribed fire and wildfire showed a significant difference in post-fire fuel loads for up to 5 years post fire based on a one-tailed
t-test with unequal variances (
t = 1.67,
p < 0.00001). Following the immediate reduction in fine-fuels one-year post fire (
n = 12,
µ = −394 kg/ha), there was a gradual increase in fine-fuels with return to pre-fire levels between 3 and 5 years (
Table A4). Second-year post-fire average reductions (
n = 11,
µ = −288 kg/ha) in estimates of fine-fuels were about 3/4 of first-year reductions. Average reductions of fine-fuels (
n = 9,
µ = −294 kg/ha) in the third year were similar to second-year reductions and 3/4 of first-year reductions, on average. Fourth-year estimates of average yearly changes in fine-fuel surpassed initial pre-fire conditions (
n = 8,
µ = 54 kg/ha) with fifth-year estimates slightly below. Individually, some fire treatment areas showed recovery of estimated fine-fuels to pre-treatment levels within 1–2 years and upwards of 5 years.
The individual effect of fires on burned area polygons varied due to treatment type, a reflection of seasonality, timing, and burning objectives. Average reductions in fine-fuels by wildfire events tended to be much greater on average than those of prescribed fires (
Table A4). Although, with a Bonferroni adjustment that shifted significant
p-values from 0.05 to 0.0125 for multiple tests (4 years), there were no statistically significant differences for prescribed fire and wildfire in post-fire years and the comparisons were hampered by low and diminishing sample size over time. When fire types (prescribed fire and wildfire) were grouped and burned areas were compared to unburned control plots, yearly differences were significant. One-tailed
t-tailed tests with Bonferroni adjustments revealed significant differences in the first (
t = −2.80,
p = 0.006) and second year (
t = −3.03,
p < 0.0001) post fire. Third (
t = −1.73,
p = 0.054), fourth (
t = −0.55,
p = 0.30), and fifth (
t = −2.67,
p = 0.01) post-fire years were not significantly different with Bonferroni adjustments in average fuel reductions due to fire; however, sample size was halved by year 5 and a greater proportion of year 3–5 post-fire samples included 2019, a year of enhanced productivity.
4. Discussion
With increases in wildfire activity in recent decades [
1,
2], there is a growing need for an accurate, updateable, and expandable model of fine-fuels. Our annual estimates of fine-fuel across a wildlife refuge in southern Arizona provide spatially explicit information for fire-risk mitigation strategies, wildlife management, and to further our understanding of the ecological outcomes of wildfire and prescribed fire in non-forested ecosystems. Determining the contemporary trends of fine-fuels and their response to fire on the refuge can help inform fuel treatments in similar systems that have undergone invasion by non-native grasses and experienced a legacy of different land uses. As expected, areas within the refuge that experienced both prescribed fire and wildfire showed substantial reduction in estimates of fine-fuels the first year following fire and recovered to pre-fire levels within 3–5 years. Our comparison with unburned control areas suggests that fuel conditions may fluctuate less widely between years than previously expected, with wildfire having a greater impact on fuel reductions (−520 kg/ha) than prescribed fire (−160 kg/ha).
Our study demonstrates that Sentinel-2A imagery can produce spatially explicit fine-fuel estimates at relatively high spectral (13 bands) and spatial resolution (10 m) over a broad spatial extent (40 km × 30 km). These improvements are necessary to assess pre- and post-fire conditions and evaluate fuel treatment efficacy and longevity. Previous efforts to map fuels in arid and semi-arid regions have typically relied on coarse to moderate resolution imagery [
42,
46,
47] or focused on generating high-resolution predictions within a narrow study region of interest [
48,
49]. The increased resolution and spatial extent of fuel estimates in our study are more suitable for fuel management planning and decision making. The relatively high resolution of our fuel estimates is particularly important for the patchy and heterogenous fine-fuel in semi-arid grasslands, which is in contrast to the more traditionally studied continuous and homogenous fine-fuels of more mesic systems [
50]. Due to the ongoing, publicly available data collection by the Sentinel-2 mission, the methodology we implemented is updateable and applicable to other ecosystems where improvements to fine-fuel monitoring is needed.
Our model estimates of fine-fuels showed positive linear agreement with validation data in both 2015 (adjusted-R
= 0.52) and 2020 (adjusted-R
= 0.63) with some sources of unexplained variation (
Figure 3). The 2020 model showed slight bias towards over predicting fine-fuel, which may be due to the acquisition of 2020 images 2 weeks later in the growing season compared to the 2015 training year. Additional unexplained variation may be attributable to a resolution mismatch between the image predicted (10 m) and field observed (2 m) estimates of fine-fuel. Nonetheless, the number of variables we included in model parameterization was far less than previous efforts [
42], which reduces overfitting of models, a potential drawback of machine learning methods that results in the inability to predict future observations reliably. Although additional variables such as vegetation and soil indices may increase model accuracy, our multi-year validation allowed for the reliable development of a time series of fine-fuel estimates. Key elements of our analysis that allowed for this time series included rigorous image corrections and radiometric normalizations to match spectral data across scenes and years. Our model results expand on other remote sensing studies that show the importance of shortwave and near-infrared bands for characterizing vegetation dynamics in grasslands and other drylands, especially where biomass is largely composed of non-photosynthetic vegetation [
51,
52]. Our models also show that considering spectral information during both peak vegetation greenness and vegetation dormancy can improve estimates of fine-fuels in grasslands and other drylands. Assessment of grassland fuels is particularly important in the context of invasive annual and perennial grasses that may show high interannual variation in fuel hazard [
53,
54]. We found that the spectral range and frequent return interval of the Sentinel-2A vehicle (10-day return interval) was compatible with yearly fine-fuel monitoring objectives and expect the Sentinel-2B vehicle to achieve similar results, offering a combined 5-day return interval of the Sentinel-2 mission.
Our fuel models can be readily employed to monitor fine-fuel changes over multiple years, which is an improvement from a previous one-year assessment [
42]. This progress is necessary to detect fine-fuel changes attributable to climate, land-use, and fire conditions. For example, our time series accommodated the large interannual climate variability of the 2015–2020 study period, which included wetter (2019) and much drier (2020) than average conditions in consecutive years. This detection of year-to-year variation can largely be attributable to conducting field validations in both wet (2015) and very dry (2020) years. Our models will be useful moving forward as 2021 is one of the wettest summers on record and will likely create high fine-fuel loads and additional wildfire risk, especially if a subsequent dry year elevates flammability. Indeed, drier conditions interspersed with periods of more intense rainfall is the forecasted trend for the southwestern U.S. [
55,
56] and is likely to create additional fire risk that our models can help inform.
Our study area is widely representative of increasingly fire prone drylands throughout the western U.S. and world [
10,
57] due to invasion by non-native grasses [
7,
17]. While we did not specifically account for the differences in non-native and native grass contributions to fine-fuels, our estimates of fuel load within 2015–2020 burn perimeters (1400–1800 kg/ha) are twice the previous estimates of Lehman lovegrass invaded semi-arid grasslands in southeastern Arizona in dry summers (900 kg/ha) and considerably higher than the fine-fuel in native semi-arid grasslands in the region (300–700 kg/ha) (Cox et al. 1990). Non-native lovegrasses, introduced to provide erosion control and forage, have proliferated over vast parts of the refuge and semi-arid grasslands in the region over the last several decades and have enhanced fire continuity and spread [
58]. Their general high tolerance to drought and livestock grazing [
59], which is widespread in areas adjacent to the refuge, indicate that they will continue to present wildfire risk in the future. Although the refuge has not experienced livestock grazing since 1985, future studies can compare how fuel load from native and non-native invasive grasses changes with grazing and other land-uses.
While our results indicate reductions in fine-fuel for up to three years due to prescribed burning, previous studies have suggested that invasive lovegrasses may increase or show no change, in relation to decreased abundance of native grasses, even when fire return intervals are frequent [
52,
60,
61]. Beginning one-year post fire, our analysis showed a regeneration of fine-fuels after prescribed fire surpassing pre-fire fine-fuel estimates in the third to fourth year post fire (
Figure 6b). Future prescribed fires can balance the need for a longer recovery time of native grasses, and the habitat requirements of masked bobwhite quail, with the need to manage fire risk. Our results generally indicate areas of patchy high fine-fuel abundance, which is consistent with the preference of Lehmann lovegrass occurring on soils of high sand and low clay content [
62] and the biophysical relationships of fire distribution of grasslands in the southwestern U.S. [
63]. These and other environmental conditions that increase the abundance of Lehmann lovegrass and associated fine-fuels can be incorporated into future burn plans [
64]. Further analysis of Sentinel-2A imagery in the manner described herein may eventually allow for biophysical-spectral mapping of Lehmann lovegrass dominated sites, allowing for greater quantification and characterization of the threat of invasive grasses.
Our study demonstrated that wildfire decreased fuel by three times the amount of prescribed fire and increased the length of time until fine-fuels approach pre-fire estimates (
Figure 6). This difference is largely because prescribed burning typically happens under the cooler and relatively wetter conditions in the spring, which reduces the risk of unwanted spread but leads to less fuel consumed than the hotter and drier conditions of summer wildfires [
65]. An exception was the 2016 Airport prescribed fire that consumed an amount of fine-fuel comparable to most wildfires on the refuge due to its ignition in June. While recent calls for more prescribed fire [
21] may be necessary in many over-crowded forests and grasslands encroached on by woody species, burn plans in drylands invaded by non-native flammable species should carefully weigh the cost of increasing fire return interval and continued spread of the non-native species. These invasions may result in unfavorable wildlife habitat and lead to land degradation.
The range of 3 to 5 years needed for recovery of fine-fuels that we found in our study is likely influenced by the seasonality and size of the burn and annual climate. For example, the slight increase in fine-fuels the year following fire in 2019 was likely due to above average precipitation, the late-winter timing of the prescribed East Gate fire, and the limited areal extent of the High Peak (3.4 ha) fire. Winter burning combined with favorable spring growing conditions promotes plant recruitment and growth, thereby enhancing fine-fuels, especially with the presence of invasive grasses. The relatively short recovery period of fine-fuels found in our study supports previous findings of semi-arid grassland fires [
16,
66,
67]. Although we expected higher interannual variability of fine-fuels associated with climatic fluctuations, the variability was relatively low and might be explained by the high resilience of grasses to fire, the dominance of perennial grasses in our study area that have lower interannual changes than annual grasses, and the maintenance of fuel through time by the highly productive invasive Lehmann lovegrass.
Our results provide a means for fire managers to prioritize strategic actions where critical natural resource values, infrastructure, and human health are at stake. Updateable, high-resolution fine-fuel maps allow for managers to move from a reactive approach to a proactive data-driven approach that can be used in an adaptive framework [
68] to better understand and manage potentially hazardous fuels through time. Additionally, these annual fuel layers can be overlaid with other data relevant to assessment of values at risk, such as critical habitat for endangered species (e.g., masked bobwhite quail at BANWR) or areas of high erosion potential, to help coordinate prescribed burning and other treatments. Our fuel models represent an important testing and calibrating opportunity for verification of simulations of fire behavior outputs [
69] that have historically not performed well under conditions of patchy and heterogenous fuels in drylands [
50].
In particular, our fine-fuel model is geared toward use in next-generation fire simulation programs such as the QUIC-Fire (Linn et al., 2020) tool for prescribed fire planning that allows for inclusion of a continuous fine-scale model of surface fuels. Our fuel models of BANWR represent an important testing and calibrating opportunity for wildfire simulations and an opportunity to implement an adaptive management approach to refuge operations, including on-going efforts to improve the population viability of the masked bobwhite quail. Past prescribed fires that have known ignition and spread patterns can be analyzed to help calibrate forecasting tools to better inform future burning outcomes based on fire weather and fine-fuel estimates. When observed versus modeled fuel and fire simulations are iteratively tested, they can be used to plan future treatment scenarios with otherwise unknowable outcomes. Our study provides an avenue for potential integration of updateable fine-fuels across dryland ecosystems more broadly, to formulate a better understanding of fire risk and fuel treatment outcomes in the western U.S. and throughout the world.