Continuous 1985–2012 Landsat Monitoring to Assess Fire Effects on Meadows in Yosemite National Park, California

Abstract

To assess how montane meadow vegetation recovered after a wildfire that occurred in Yosemite National Park, CA in 1996, Google Earth Engine image processing was applied to leverage the entire Landsat Thematic Mapper archive from 1985 to 2012. Vegetation greenness (normalized difference vegetation index (NDVI)) was summarized every 16 days across the 28-year Landsat time series for 26 meadows. Disturbance event detection was hindered by the subtle influence of low-severity fire on meadow vegetation. A hard break (August 1996) was identified corresponding to the Ackerson Fire, and monthly composites were used to compare NDVI values and NDVI trends within burned and unburned meadows before, immediately after, and continuously for more than a decade following the fire date. Results indicate that NDVI values were significantly lower at 95% confidence level for burned meadows following the fire date, yet not significantly lower at 95% confidence level in the unburned meadows. Burned meadows continued to exhibit lower monthly NDVI in the dormant season through 2012. Over the entire monitoring period, the negative-trending, dormant season NDVI slopes in the burned meadows were also significantly lower than unburned meadows at 90% confidence level. Lower than average NDVI values and slopes in the dormant season compared to unburned meadows, coupled with photographic evidence, strongly suggest that evergreen vegetation was removed from the periphery of some meadows after the fire. These analyses provide insight into how satellite imagery can be used to monitor low-severity fire effects on meadow vegetation.

1. Introduction

The evaluation of vegetation change over time can be conducted in a variety of ways, from traditional field observations to monitoring with terrestrial or space borne Earth Observation Systems (EOS) [1]. Satellite data provide repeated measurements, but the length of the temporal record, pixel and spectral resolution, and the revisit rate dictate the type of surface processes that can be analyzed [2]. While varied in temporal, spatial, and spectral resolutions, satellite EOS provide tools to evaluate vegetation attributes over time [3,4]. Vegetation estimates from moderate/low resolution MODIS (250/500 m) and AVHRR (1.1 km) sensors have been extensively and successfully used to measure and interpret vegetation dynamics [5,6].

Additionally, recent advances in computational process times and access to free Landsat imagery have increased the applicability of mid-resolution (i.e., Landsat 30 m) time series to detect disturbance events, and monitor short-and long-term land-use/land-cover trends [2,7,8]. The advent of client-thin EOS processing through Google Earth Engine (GEE) has proven effective to circumvent many of these previous constraints [9,10,11]. Due to the 40+ year temporal record provided by the Landsat satellite program, finer resolution image data are increasingly being applied to investigate the temporal response of vegetation to climate and land use changes [12]. While limited to a 16-day revisit rate, Landsat imagery has been used to study the impact of isolated disturbance events over a short duration and longer term processes associated with climate patterns [13]. The spatial resolution of Landsat also provides the benefit of evaluating small-scale features on the landscape [14].

Landsat imagery has been proven to be especially effective in evaluating forest characteristics over time [15,16], with measurements regularly applied to identify abrupt forest changes [17]. Forest disturbance datasets derived from Landsat include LANDFIRE Vegetation Disturbance [18,19], the North American Carbon Program Vegetation Change Tracker [20], and Web-enabled Landsat Data [21,22]. Most recently, the Global Forest Change effort employed GEE processing to detect global forest loss continuously using Landsat imagery [9].

While Landsat satellite imagery provides consistent, objective spectral measurements over space and time well-suited to detect abrupt forest disturbance, detecting disturbances in non-forested landscapes remains a challenge. Further, algorithms have difficulty discerning subtle effects or understory changes in Landsat imagery [23]. Although algorithm refinements continue to better separate subtle changes and background noise [19,24,25], current disturbance products are insufficient for identifying disturbance events and studying disturbance effects in systems like mountain meadows, where vegetation composition is either mixed or largely characterized by herbaceous vegetation. Monitoring the occurrence and effect of disturbances (extreme weather events, wildfire, land use, etc.) within mountain meadows using EOS has only been applied in isolated cases [26]. Mountain meadows can be monitored using ground-based methods [27], yet may be difficult to access given their often remote locations. Further, the relatively small size and lack of potentially obscuring canopy cover make them excellent candidates for monitoring with finer resolution satellite data, such as Landsat [26].

Montane meadows provide a wide variety of ecosystem services (e.g., nutrient cycling, biodiversity, carbon storage), and their role in regulating floods, filtering sediments and providing clean water is particularly important in places like the Sierra Nevada, where montane watersheds provide nearly all the clean water needed to support agriculture and municipal water needs [28]. It is therefore critical to have timely information regarding how meadows respond to climate change, wildfire and fire suppression, and other human uses [29,30,31]. Fire can influence meadow vegetation in numerous ways, depending on the timing, location, extent, and severity of the fire, the vegetation surrounding the meadow, and the water availability within the meadow [32]. Indirect effects of fire may include changes in watershed processes that influence short- or long-term alterations in sediment and flow regimes within the meadow. Alternatively, meadows may be directly affected by fire in a variety of ways, with low-severity fire potentially affecting peripheral vegetation while only minimally affecting vegetation within the meadow, or high-severity fire and/or drought conditions potentially altering soils and vegetation composition or productivity, resulting in long-term changes to the meadow [29,32].

The magnitude of vegetation change associated with a disturbance event will affect the ability of statistical filtering to effectively detect breaks in an EOS time series [33,34]. Products from the Monitoring Trends in Burn Severity effort or LandTrendr can help identify subtle disturbances, such as low severity fires in herbaceous meadows [35,36], but are not designed to monitor long-term, post-fire effects. To evaluate how meadow vegetation greenness responds to wildfire, including meadows not flagged by statistical detection processes, MTBS data was used to identify the presence of fire events across the Sierra Nevada. GEE image processing was applied to derive vegetation index values (normalized difference vegetation index (NDVI)) across the entire Landsat 5 Thematic Mapper (TM) archive for 1985–2012 for a subset of meadows in the Sierra Nevada mountain range of California with similar biophysical characteristics.

NDVI = (TM_band4 − TM_band3)/(TM_band4 + TM_band3)

(1)

NDVI values and trends were evaluated seasonally across burned and unburned meadows to determine how vegetation greenness changed temporally in response to fire relative to similar unburned sites. Measuring the response of meadow vegetation to a single wildfire event provides an opportunity to begin to identify potential ecological benefits or drawbacks caused by wildfire on the landscape. We focus here on an isolated disturbance event and a subset of protected meadows in Yosemite National Park as a demonstration of an approach that could be extended to examine meadow responses to disturbance or other environmental factors.

2. Materials and Methods

2.1. Description of Sierra Nevada Meadows

To assess the effect of fire on vegetation in Sierra Nevada meadows, we identified twenty-six subalpine meadows (>2 ha) in Yosemite National Park [37] that had either burned or were adjacent to the Ackerson Fire of 1996 (Figure 1; Supplementary S1). All meadows are located in the Upper Middle Tuolumne River sub-watershed. These meadows are largely unaffected by recent anthropogenic disturbances through federal protections, and are characterized by herbaceous vegetation, typically mixtures of grass (Poaceae), sedge (Cyperaceae), rush (Juncaceae) and interspersed woody (Salix spp.) riparian vegetation [29], that cover inceptisol soils and granodiorite or glacial drift geology. According to the Yosemite National Park Vegetation Classification and Mapping Project [38], woodlands on the periphery of the meadows are primary composed of California Red Fir (Abies magnifica), Western White Pine (Pinus monticola), and Sierra Lodgepole Pine (Pinus contorta var. murrayana). A visual assessment of meadow conditions using contemporary aerial photography spanning 1993–2014 suggests that all meadows either have a single riparian channel or a subsurface drainway, and no evidence of prominent human influences (corrals, buildings, hiking trails, grazing). The mean elevation, average precipitation, and average temperature differ across the burned and unburned meadows (p < 0.05 for all three variables), with a slightly lower altitude, less rainfall, and warmer temperatures in burned versus unburned locations.

While the 26 meadows in the Upper Middle Tuolumne River sub-watershed share many characteristics, the meadows in the northwestern part of the study area are distinguished by the presence of fire in late 1996. A “changepoint” analysis in R Stats [39,40], intended to identify deviations in mean and/or variance in the NDVI time series to find single or multiple changepoints within data, only identified a break in the NDVI time series for 6 meadows coinciding with the fire event. According to the California Department of Forestry Fire and Resource Assessment Program [39] and Monitoring Trends in Burn Severity database [35,41], the Ackerson Fire contributed to low-to-moderate severity fire in Yosemite National Park starting in August 1996. While 16 of the 26 meadows intersect the fire footprint, only 15 meadows register a change based on Differenced Normalized Burn Ratio (dNBR) in imagery collected before and after the fire event.

NBR = (TM_band4 − TM_band7)/(TM_band4 + TM_band7)

(2)

dNBR = NBR_pre-fire − NBR_post-fire

(3)

A visual comparison of aerial photographs from 1993 (USGS National Aerial Photography Program) and 2005 (USDA National Agriculture Imagery Program) confirm the observable loss of vegetation within or along the periphery of 14 of 16 meadows in the fire footprint to further illustrate the influence of fire (Figure 2). All change point and MTBS fire metrics are summarized in Supplementary S1. The most obvious change is the loss of tree cover in areas immediately surrounding each of the 14 burned meadows. The remaining 12 meadows were included in the analysis to evaluate undisturbed sites with similar biophysical characteristics.

2.2. Google Earth Engine (GEE) Image Processing and Noise Removal

GEE image processing was applied to evaluate vegetation greenness (NDVI) patterns using the Landsat 5 Thematic Mapper archive spanning 1985–2012. Landsat scenes were co-registered, normalized to at-surface reflectance, and atmospherically corrected [43]. Landsat 7 imagery was excluded from this study, and cross-sensor calibration was not required. NDVI values calculated at a 30-meter pixel resolution were spatially averaged and summarized for the 26 meadows from imagery compiled every 16 days over a 28-year period (over 600 observations per meadow).

Zonal averaging by meadow footprint was applied to address potential horizontal pixel shifts remaining in co-registered images. Data quality assessment was implemented using a combination of GEE computed Automatic Cloud Cover Assessment (ACCA) cloud masks [44], albedo measurements [43], and NDVI outliers by applying a z-score to tabular outputs from GEE (more than two standard deviations above or below the monthly mean), effectively removing discrete measurements likely resulting from snow, clouds, and shadows from the time series analysis. Any remaining NDVI anomalies were manually examined in Landsat imagery. Monthly NDVI composites were assembled using the filtered NDVI time series to create a consistent time series for each meadow in the study (maximum of 336 monthly NDVI values with an average of 15 monthly values removed from each time series due to contamination; Supplementary S2).

z-score = (x − mean)/sd

(4)

2.3. Testing for Differences in NDVI Mean

To evaluate the difference between burned and unburned meadows, we calculated pre-fire date NDVI mean, post-fire date NDVI mean, and post-fire NDVI outliers using the entire monthly time series. A hard break (August 1996) was defined for all meadows to evaluate fire response in burned meadows relative to unburned meadows. The pre-fire NDVI mean (1 January 1985 to 7 January 1996) was calculated for all burned and unburned meadows in growing (April–September) and dormant (October–March) seasons, and compared to post-fire NDVI means to measure recovery time. Post-fire NDVI means were calculated for the 16-year, post-fire time series (11 January 1996 to 12 January 2012) (Figure 3), as well as shorter periods spanning 1 to 6 years to evaluate potential short-term recovery periods. Any month where the fire may have been active was excluded from the analysis.

2.4. NDVI Outlier Frequency

NDVI outlier count was calculated by comparing the monthly NDVI values after the fire date (196 total months) to the pre-fire NDVI mean (±1 sd) for the same month compiled using a maximum of 139 pre-disturbance records for each meadow. A count was updated for every monthly NDVI value that fell within one standard deviation of pre-fire mean (±1 sd) and below pre-fire mean (−1 sd). To evaluate potential recovery effects, frequency was compared between burned and unburned meadows for the 5 years immediately following the fire, 5 years after the fire, 10 years after the fire, and for the entire post-fire time series. Finally, NDVI outlier counts were calculated for both the growing and dormant seasons of each analysis period to identify possible vegetation composition changes in the burned meadows, notably changes from evergreen trees to herbaceous vegetation types that have distinctive phenology between seasons [45].

2.5. NDVI Trends

To evaluate the difference between burned and unburned meadows, NDVI trends were calculated and compared using non-parametric Mann Kendall [46,47] and Theil Sen [48,49] tests for monotonic trends for the full time series (1 January 1985–12 January 2012). Monotonic tests are best suited for unbroken time series. While the low burn severity and limitations in flagging change-points suggest that the fire event may not need to be removed from monotonic tests, we also ran pre-fire trend tests (1 January 1985–7 January 1996) and post-fire time trend tests (1 January 1997 to 12 January 2012). All slope analyses separated dormant and growing seasons using a deseasonalized Mann Kendall test for monotonic trends implemented in the R package “kendall” [50]. Winter precipitation records were regressed against NDVI, and confirm that NDVI may be sensitive to exogenous precipitation influences in only 5 of the 26 meadows included in the study (r² > 0.5). An additional requirement of the Mann Kendall trend test is that observations are independent, since positive autocorrelation increases the probability of incorrectly rejecting the null hypothesis of no trend [51]. We examined mean NDVI values from 1985–2012 in growing and dormant seasons for evidence of temporal autocorrelation. Meadows that showed evidence of serially dependent data were subjected to a de-trended analysis implemented in the R package “zyp” [52,53]. Mann Kendall tests for trend direction (tau) and trend significance (p ≤ 0.10) were coupled with NDVI linear trend estimates, calculated using Theil Sen median slope fitting executed in the R package “mblm” [48,54].

3. Results

3.1. Mean NDVI—Before and After Fire

A comparison between pre-and-post-fire NDVI suggests that burned meadows exhibit a statistically lower NDVI mean after the Ackerson Fire. Lower NDVI is evident across all dormant season comparisons regardless of the interval length (

Table 1. Growing and dormant season mean NDVI for pre-fire date and post-fire date time series for burned meadows. Statistical difference based on one-tail t-test assuming unequal variances (95% significant values in bold).

Burned	Post Fire Interval Length			Pre Fire Mean	Post Fire Mean	Pre–Post	P(T ≤ t)
Dormant	January 1996	to	January 1997	0.445	0.342	0.103	0.017
Dormant	January 1997	to	January 1998	0.445	0.247	0.197	<0.001
Dormant	January 1998	to	January 1999	0.445	0.307	0.137	0.003
Dormant	January 1999	to	January 2000	0.445	0.375	0.070	0.056
Dormant	January 2000	to	January 2001	0.445	0.310	0.135	0.003
Dormant	January 1996	to	January 2001	0.445	0.318	0.126	0.004
Dormant	January 2001	to	January 2006	0.445	0.292	0.153	0.001
Dormant	January 2006	to	January 2012	0.445	0.300	0.145	0.001
Dormant	January 1996	to	January 2012	0.445	0.318	0.126	0.004
Growing	January 1996	to	January 1997	0.545	0.585	−0.039	0.088
Growing	January 1997	to	January 1998	0.545	0.444	0.101	0.001
Growing	January 1998	to	January 1999	0.545	0.529	0.017	0.262
Growing	January 1999	to	January 2000	0.545	0.551	−0.005	0.420
Growing	January 2000	to	January 2001	0.545	0.538	0.007	0.386
Growing	January 1996	to	January 2001	0.545	0.524	0.021	0.205
Growing	January 2001	to	January 2006	0.545	0.519	0.026	0.149
Growing	January 2006	to	January 2012	0.545	0.515	0.030	0.118
Growing	January 1996	to	January 2012	0.545	0.521	0.025	0.169
All	January 1996	to	January 2012	0.500	0.419	0.080	0.008

). Comparisons of NDVI means over the growing season show no statistically significant differences in greenness between pre- and post- fire for all periods except for 1997–1998. Unburned meadows were also tested using the same procedure to assess differences in NDVI means over the dormant and growing seasons (

Table 2. Growing and dormant season mean NDVI for pre-fire date and post-fire date time series for unburned meadows. Statistical difference based on one-tail t-test assuming unequal variances (95% significant values in bold).

Unburned	Post Fire Interval Length			Pre Fire Mean	Post Fire Mean	Pre–Post	P(T ≤ t)
Dormant	January 1996	to	January 1997	0.349	0.288	0.103	0.108
Dormant	January 1997	to	January 1998	0.349	0.241	0.197	0.012
Dormant	January 1998	to	January 1999	0.349	0.284	0.137	0.092
Dormant	January 1999	to	January 2000	0.349	0.299	0.070	0.157
Dormant	January 2000	to	January 2001	0.349	0.301	0.135	0.164
Dormant	January 1996	to	January 2001	0.349	0.294	0.126	0.125
Dormant	January 2001	to	January 2006	0.349	0.291	0.153	0.104
Dormant	January 2006	to	January 2012	0.349	0.306	0.145	0.178
Dormant	January 1996	to	January 2012	0.349	0.302	0.126	0.160
Growing	January 1996	to	January 1997	0.492	0.544	-0.039	0.030
Growing	January 1997	to	January 1998	0.492	0.368	0.101	<0.001
Growing	January 1998	to	January 1999	0.492	0.493	0.017	0.472
Growing	January 1999	to	January 2000	0.492	0.507	-0.005	0.267
Growing	January 2000	to	January 2001	0.492	0.516	0.007	0.135
Growing	January 1996	to	January 2001	0.492	0.484	0.021	0.361
Growing	January 2001	to	January 2006	0.492	0.481	0.026	0.322
Growing	January 2006	to	January 2012	0.492	0.482	0.030	0.331
Growing	January 1996	to	January 2012	0.492	0.483	0.025	0.350
All	January 1996	to	January 2012	0.427	0.392	0.080	0.141

). While most dormant and growing season comparisons in the unburned, control group show no statistically significant difference, a significantly lower growing season NDVI was also observed in unburned meadows in the 1997–1998 period similar to the burned meadows. Significant differences observed in all meadows during 1997–1998 were not caused by fire. A visual assessment of Landsat TM imagery during June 1998 shows standing water in both unburned and burned meadows, which is likely a result of the heavy precipitation in the 1997/1998 winter season.

3.2. NDVI Outlier Frequency

A second series of tests compared monthly NDVI values after the fire date (from September 1996) to a monthly NDVI mean compiled for each meadow using 11 years of NDVI records preceding the fire. To test how often a meadow intersected the historical range in greenness (mean − 1 sd ≤ NDVImonth ≤ mean + 1 sd) or fell below the pre-fire NDVI mean (NDVImonth < (mean − 1 sd)), we tallied each month each of these thresholds were met across varied 5 to 16 year periods (

Table 3. ANOVA single factor test of statistical difference between 14 burned and 12 unburned meadows indicates burned meadows are less likely to fall within one standard deviation of pre-fire mean (mean ± 1 sd) and more likely to fall below pre-fire mean (mean − 1 sd) than unburned meadows based on evaluating monthly NDVI values across 5 to 16 year periods following the disturbance. Frequency percentage based on mean outlier count relative to the number of months evaluated.

# of Months	Period	Burned (Mean)	Frequency (%)	Unburned (Mean)	Frequency (%)	p-Value
Within Range	1996–2001	35.85	60%	42.42	71%	0.002
Within Range	2001–2006	33.71	56%	39.42	66%	0.03
Within Range	2006–2012	41.07	54%	48.75	64%	0.004
Within Range	1996–2012	108.21	55%	128.75	66%	0.003
Within Range	2001–2012	74.14	55%	87.58	64%	0.007
Below Range	1996–2001	21.14	35%	11.42	19%	<0.001
Below Range	2001–2006	24.86	41%	11.50	19%	<0.001
Below Range	2006–2012	31.29	41%	13.58	18%	<0.001
Below Range	1996–2012	78.07	40%	36.50	19%	<0.001
Below Range	2001–2012	55.77	41%	24.92	18%	<0.001

).

Unburned meadows had monthly NDVI values fall within the pre-fire range (64%–71%) more often than burned meadows (54%–60%). Moreover, unburned meadows had monthly NDVI values fall below the pre-fire range in 18–19 percent of the months following fire while burned meadows had lower than normal NDVI values in over 35–41 percent of the post-fire months. Tests segmenting NDVI outlier counts by growing and dormant seasons further illustrate that burned meadows are less likely to fall within one standard deviation of pre-fire mean (mean ± 1 sd) and more likely to fall below pre-fire mean (mean − 1 sd) than unburned meadows regardless of season. Unburned meadows are within the pre-fire NDVI range with the same frequency between growing season (67% of months) and dormant season (64%) for the period 1996–2012. Burned meadows have lower frequency (42%) within the pre-fire range in the dormant season but the frequencies are similar to those in unburned meadows during the growing season (69%).

3.3. NDVI Trends

Across the entire time series (1985–2012), 17 of 26 meadows had statistically significant negative Kendall Tau/Sen trends (p ≤ 0.10) in the dormant season, and 16 of 26 meadows had statistically significant negative trends in the growing season.

Table 4. Statistically significant, long term (1985–2012) monotonic trends reported in 17 meadows over the dormant season and 16 meadows in the growing season. * indicates that winter precipitation may exert an exogenous influence on NDVI trend (r² > 0.5).

Dormant Season 1985–2012					Growing Season 1985–2012
	Kendall		Theil Sen			Kendall		Theil Sen
ID	tau	p-value	Slope	MTBS dNBR	ID	tau	p-value	Slope	MTBS dNBR
2807	−0.233	0.09	−0.0022	0.0	2807	−0.270	0.05	−0.0016	0.0
2811	−0.497	0.00	−0.0090	82.4	2811	−0.376	0.01	−0.0027	82.4
2835	−0.550	0.00	−0.0103	60.6	2835	−0.444	0.00	−0.0038	60.6
2878 *	−0.429	0.00	−0.0067	87.1	2867	−0.259	0.06	−0.0019	0.0
2880	−0.577	0.00	−0.0113	86.0	2878 *	−0.360	0.01	−0.0029	87.1
2882	−0.259	0.06	−0.0018	0.0	2880	−0.259	0.06	−0.0018	86.0
2886	−0.434	0.00	−0.0054	0.0	2886	−0.296	0.03	−0.0016	0.0
2887	−0.667	0.00	−0.0221	205.6	2887	−0.481	0.00	−0.0045	205.6
2893	−0.503	0.00	−0.0063	0.0	2893	−0.333	0.01	−0.0029	0.0
2900 *	−0.582	0.00	−0.0096	190.8	2895	−0.212	0.12	−0.0016	70.2
2905	−0.460	0.00	−0.0069	93.5	2900*	−0.344	0.01	−0.0029	190.8
2918	−0.651	0.00	−0.0139	117.7	2905	−0.275	0.04	−0.0013	93.5
2923 *	0.222	0.10	0.0020	42.6	2918	−0.413	0.00	−0.0031	117.7
2944	−0.365	0.01	−0.0051	71.3	2944	−0.307	0.02	−0.0019	71.3
2950	−0.291	0.03	−0.0037	76.9	2959	−0.254	0.06	−0.0021	135.8
2959	−0.667	0.00	−0.0163	135.8	2971	−0.222	0.10	−0.0021	97.9
2971	−0.540	0.00	−0.0100	97.9	-

shows that many of the meadows with statistically significant trends in the dormant season were also significant in the growing season, including only four unburned meadows reoccurring on each list. Far fewer statistically significant trends were measured in pre-fire and post-fire time series, although many of the burned meadows also exhibit a similar, declining greenness in the period spanning 1997–2012 (full Mann Kendall results are available in Supplementary S3).

NDVI trends indicate that both unburned and burned meadows in the study area have become less green over time. In the dormant season, burned meadow slopes are more negative (slope = −0.012x) than unburned meadow slopes (slope = −0.003x), significant at 90% confidence level based on an ANOVA single factor test of statistical difference (p = 0.096). Burned and unburned meadow slopes are not statistically different over the growing season (p = 0.286). Finally, dormant season Sen slope (p ≤ 0.1) were regressed against MTBS differenced Normalized Burn Ratio. Results suggests that slope decreases as a function of burn severity (r² = 0.58), with the meadows with more intense burning exhibiting the steepest negative slope from 1985–2012 (Figure 4).

4. Discussion

The assessment of wildfire effects on subalpine meadows within the 1996 Ackerson Fire perimeter relative to unburned meadows outside the fire perimeter indicates that fire effects were subtle. Mean NDVI values were statistically lower for burned meadows following the fire, with long lasting differences in the dormant season and no detectable effect over the growing season. The comparison of NDVI means and outliers between the pre-fire and post-fire time series suggests post-fire effects on vegetation persist in the burned meadows, as dormant season NDVI values remain lower for the remainder of the study period. Dormant season NDVI trends further illustrate the difference between long-term meadow greenness between burned and unburned sites. The statistically significant link between higher burn severity and declining greenness, when presented alongside other dormant season NDVI effects, strongly suggests that the fire severity led to a long-term reduction in NDVI greenness in burned meadows.

The work outlined in this paper does not attempt to link the change in NDVI to an estimate of lost biomass or vegetation productivity associated with the fire event [55], instead linking a decline in greenness to loss of evergreen vegetation that would otherwise amplify NDVI in dormant months when herbaceous vegetation senesces. The primary effect of the fire that was detected in our analysis appears to be tree mortality on the edges of some meadows. These changes are signified by lower dormant season NDVI and declining dormant season trends in burned meadows, as well as photographic evidence showing that conifers adjacent to meadows were removed during the Ackerson Fire. However, monthly NDVI growing season greenness after the fire date was not significantly different from pre-fire greenness in the burned meadows, suggesting that herbaceous vegetation growth in the summer months may remain unaffected.

The invasion of subalpine meadows by conifers in the western United States is well documented [56]. Seed dispersal of herbaceous meadow species is inhibited by competition from forest species [57]. To counter the potential consequences of conifer encroachment, the mechanical removal of conifers has been shown to be an effective treatment for restoring meadow species such as aspen (Populus tremuloides) [58]. The season-specific analyses included in this study show that low-severity fire may also positively influence meadows. The removal of woody vegetation along the periphery of the burned meadows, coupled with the possible resilience of wet meadow vegetation to low-severity fire [59], suggests that fire may also serve to maintain meadow boundaries while preserving herbaceous meadow vegetation [29,60].

This application illustrates how Google Earth Engine processing may be employed to conduct near-continuous vegetation monitoring with Landsat imagery, ultimately allowing the long-term effects of disturbances to be studied more effectively. Alternative methods employing wider temporal snapshots (i.e., annual, decadal) may be able to identify image dates coinciding with disturbance events, yet may miss seasonal patterns that indicate recovery or vegetation response, particularly if effects are subtle or isolated in specific months in the phenological cycle (i.e., native vs. non-native plant phenology). In this case, NDVI values from winter and summer seasons provided important evidence that suggested post-fire changes in woody cover on the periphery of the meadows was the primary effect of the fire. Shorter periods of analysis may limit understanding regarding the timeline of fire recovery and inter-annual vegetation dynamics. The benefits afforded from using the long term archive of Landsat TM data include comparing 12 years of pre-fire NDVI records to 16 years of post-fire NDVI records, evaluating vegetation response well over a decade after a disturbance event.

The richness of the NDVI time series also allows for an analysis of intra-annual vegetation dynamics to evaluate the timing of vegetation green-up within the longer period and to distinguish between growing and dormant season patterns as a way to separate the type of vegetation within-and-directly-adjacent- to meadows contributing to NDVI values in disturbed versus controlled sites. For example, decreases in dormant season NDVI following the fire signaled the removal of evergreen vegetation. An increasing dormant season NDVI signal may indicate expansion of evergreen vegetation into meadow boundaries over time, providing an important opportunity for managers to identify where and at what rate this is occurring and to potentially identify contributing factors, which are currently unknown [61]. Collectively, these analyses can begin to describe temporal changes in montane meadows associated with disturbance and other environmental change.

While the analysis here presents a convincing ecological explanation regarding how meadows may change due to wildfire, we do not rule out other explanatory factors (e.g., climate) that may contribute to changes in NDVI over time. Future work can incorporate climate time series to include more possible co-variates. Future research may also be coupled with past research on tree encroachment [62] to investigate how wildfire may affect meadows differently across the Sierra Nevada. Understanding how wildfire operates to remove encroaching conifers along the edge of meadows and observing how meadows respond to fire for an extended period of time after a fire has occurred may ultimately help inform targeted management decisions, such as prescribed burning, based on site-specific objectives [63].

5. Conclusions

In this study, we learned that NDVI values were significantly lower (p < 0.05) for burned meadows following the fire date. Further, by having a near-continuous Landsat record to evaluate seasonal difference we determine that burned meadows consistently exhibited lower monthly NDVI in the dormant season more than a decade following fire (p < 0.05) and a more prominent trend towards declining greenness overall (p < 0.10). The comparative seasonal analyses, made possible by Google Earth Engine (GEE) cloud computing, helped corroborate aerial photographic evidence of evergreen vegetation burned along the edge of some meadows, and serves to complement former research suggesting that fire can help maintain meadow boundaries.

This study highlights the strength of applying GEE image processing to analyze Landsat TM imagery continuously from 1985–2012, as opposed to looking at only few images in 1985–2012. Had we employed annual snapshots or decadal snapshots over the growing season instead, it is very possible that we would have missed the opportunity to identify dormant season fire effects altogether, let alone the seasonal NDVI patterns that allowed us to separate deciduous meadow vegetation from evergreen conifers along meadow boundaries.

While the decision to focus on the period 1985–2012 was rooted in evaluating fine-scale vegetation changes consistently across the Landsat 5 TM 30 m archive, future GEE cloud computing applications can easily extend studies back to 1972 using Landsat Multispectral Scanner (MSS) imagery. In the near future, including new satellites such as Sentinel in the GEE archive will add to the temporal richness of imagery and allow scientists to better understand the onset of land-cover change events (e.g., fire) on the landscape. In addition to the inclusion of more imagery available in the GEE interface, the growing community of GEE practitioners will also lead to further advances in tools and techniques designed to expedite image processing and time series analysis. In this study, GEE was primarily applied in an image processing capacity, while analyses were executed external to GEE in R Stats. In the near future, processing and analysis will increasingly run in parallel in GEE, making it more similar to comprehensive processes, such as LandTrendr [36].