High-severity wildfires have been shown to have long-term impacts on freshwater ecosystems; as nutrients are mobilized, runoff and erosion can increase, and soil properties may be modified [1
]. While there is a growing literature for the effects of fire on upland vegetation types [2
], the existing information on vegetation removal by burning remains limited for most freshwater plant communities globally.
To extend this knowledge base, satellite remote sensing can be used to effectively monitor changes in high-latitude (boreal and tundra) wetland vegetation cover and productivity, especially following disturbance events such as wildfires [3
]. Most of these remote sensing studies have been carried out for non-wetland (interior boreal forest and upland tundra) vegetation cover types. Nonetheless, Potter et al. [6
] reported that the wetland tundra areas of Alaska that burned since the year 1980 had a 3:2 ratio coverage of significant positive versus negative vegetation greening trends between 2000–2010, whereas non-wetland tundra areas that burned since 1980 had a 2:5 coverage ratio of significant positive versus negative vegetation greening trends between 2000–2010. This result suggested that the wetland areas of Alaska can recover more completely and rapidly in greenness cover from recent wildfires than non-wetland land cover types; however, this supposition remains to be tested region-wide over longer time periods.
Over the past several decades, there has been an increase in the frequency and severity of boreal region wildfires in Alaska [8
]. During the 2000s, an average of 767,000 ha per year were burned statewide, which is 50% higher than in any previous decade since the 1940s. In the extreme wildfire year of 2015, nearly 60% of Alaska’s burned area was consumed at moderate-to-high severity levels [9
Most of the wildfires in the spruce forest ecosystems of Alaska are either crown or ground fires with a high enough severity to kill over-story trees [10
]. Usually, some of the organic layer of the forest floor remains, but fires in late summer following exceptionally dry or windy conditions may consume all of the organic layer, exposing mineral soil [13
]. Jiang et al. [14
] and Brown et al. [15
] reported that the post-fire thickness of the soil organic layer and its impact on soil thermal conductivity was the most important factor determining post-fire soil temperatures and thaw depth. In moderately burned sites, the presence of permafrost can mitigate the loss of the insulating soil organic layer, decrease soil drying, and increase surface water pooling.
The objective of this study was to analyze the vegetation recovery patterns of all of the Alaska wetlands that have burned at high severity since the year 2000 using a combination of the Landsat and MODIS (Moderate Resolution Imaging Spectroradiometer) satellite datasets. A statistical analysis of the changes in the MODIS vegetation index time series was conducted using the “Breaks for Additive Seasonal and Trend” method (BFAST, Verbesselt et al. [16
]). de Jong et al. [18
] analyzed trends in the normalized difference vegetation index (NDVI) satellite time series using the BFAST procedure, and detected both abrupt and gradual changes in large parts of the world, especially in shrubland and grassland biomes where abrupt greening was often followed by gradual browning.
This study was undertaken as a contribution to the NASA Arctic Boreal Vulnerability Experiment (ABoVE) field campaign, chiefly to better understand changes in related hydrologic and biogeochemical mechanisms in the years following high-latitude wildfires. One of the major questions being addressed by ABoVE, and in this type of Landsat/MODIS study, is “What processes are controlling changes in boreal–Arctic land cover properties, and what are the impacts of these changes?”
2. Materials and Methods
2.1. Landsat Burn Severity Classes
Digital maps of burn severity classes at 30-m spatial resolution were obtained from the Monitoring Trends in Burn Severity (MTBS; www.mtbs.gov
) project, which has consistently mapped fires greater than 1000 acres (405 ha) across the United States from 1984 to the present [19
]. The MTBS project is conducted through a partnership between the United States (US) Geological Survey (USGS) National Center for Earth Resources Observation and Science (EROS) and the Unites States Department of Agriculture (USDA) Forest Service. Landsat data have been analyzed through a standardized and consistent methodology by the MTBS project.
The normalized burn ratio (NBR) index was calculated by MTBS using approximately one-year pre-fire and post-fire images from the near infrared (NIR) and shortwave infrared (SWIR) bands of the Landsat sensors, with reflectance values scaled to between 0–10,000 NBR units.
NBR = (NIR − SWIR)/(NIR + SWIR)
Pre-fire and post-fire NBR images were next differenced for each Landsat scene pair to generate the Relative dNBR (RdNBR) [20
RdNBR = [(NBRpre-fire − NBRpost-fire)]/√ABS (NBRpre-fire)
The RdNBR severity classes of low, moderate, and high (LBS, MBS, HBS) have been defined previously by Miller and Thode [20
] and cover a range of −500 to +1200 over burned land surfaces. Positive RdNBR values represent a decrease in vegetation cover and a higher burn severity, while negative values would represent an increase in live vegetation cover following the fire event.
2.2. MODIS Vegetation Index Time Series
NASA’s MODIS (Moderate Resolution Imaging Spectroradiometer) satellite sensors Terra and Aqua have been used to generate a 250-m resolution NDVI (MOD13) global product on 16-day intervals since the year 2000 [21
]. The MODIS Collection 6 NDVI data set provides consistent spatial and temporal profiles of vegetation canopy greenness according to the equation:
NDVI = (NIR − Red)/(NIR + Red)
where NIR is the reflectance of wavelengths from 0.7 μm to 1.0 μm, and Red is the reflectance from 0.6 μm to 0.7 μm, with values scaled to between 0 and 10,000 NDVI units to preserve decimal places in integer file storage. Low values of NDVI (near 0) indicate barren land cover, whereas high values of NDVI (above 8000) indicate dense canopy greenness cover.
The MOD13 250-m vegetation indices (VIs) have been retrieved from daily, atmosphere-corrected, bidirectional surface reflectance. The VIs were computed from MODIS-specific compositing methods based on product quality assurance metrics to remove all of the low-quality pixels from the final NDVI value reported. Cloud and water pixels were identified and excluded using other MODIS atmospheric data masks. From the remaining good-quality NDVI values, a constrained view-angle approach (closest to nadir) then selected the optimal pixel value to represent each 16-day compositing period. These MOD13 data sets were downloaded from the files that were available at modis.gsfc.nasa.gov/data/dataprod/mod13.php for time series analysis across Alaska wetland locations.
2.3. Elevation and Land Cover Map Layers
Digital elevation (in vertical meters) for Alaska was derived from USGS [23
] mapping at 300-m ground resolution. Wetland cover was mapped for the state at 30-m ground resolution from the 2011 National Land Cover Dataset (NLCD) of Alaska ([24
]; available at www.mrlc.gov/nlcd11_leg.php
). The overall thematic accuracy for the previous Alaska NLCD was 76% at Level II (12 classes evaluated). For contextual comparison purposes, the open water (class 11), barren land (class 31), and evergreen forest (class 42) classes of this NLCD were mapped with high user’
s accuracy, while the herbaceous wetland (class 95) was mapped with moderate user’
For this study, the NLCD woody wetland (class 90) together with all of the herbaceous wetland pixels were combined into one class, and were all overlaid 200 × 200-m resolution areas with a majority of the wetland surface coverage identified and mapped for the entire state (Figure 1
). This combined wetland coverage was overlaid with statewide MTBS high burn severity (HBS) class pixels from the years 2001 to 2015, and with MODIS 250-m summer season NDVI (from the composite Julian day 177; 26 June) images for each of these years to carry out a time trend analysis of the burned wetland area NDVI changes statewide. “Pre-fire” MODIS NDVI values were all derived from the Julian day 177 NDVI from the year before the fire date for change detection.
The section of the Julian date 177 for NDVI change detection over time was not an arbitrary choice, but rather was determined to be a seasonally consistent metric of green cover change, since 26 June is nearly always near the seasonal maximum in interior Alaska for green cover, which was verified by examining thousands of pixels in time-stacked NDVI maps of Alaska wetland locations. Wetland areas that covered less than a majority of 200 × 200-m resolution areas in the statewide grid were too small to be matched consistently with MODIS 250-m summer season NDVI, and were therefore not included in the results.
2.4. Statistical Analysis Methods
The BFAST (Breaks for Additive Seasonal and Trend) methodology was applied to a MODIS NDVI monthly time series for selected wetland locations that covered the majority of a 250 × 250 m pixel area within severely burned locations. BFAST was developed by Verbesselt et al. [16
] for detecting and characterizing abrupt changes within a time series, while also adjusting for regular seasonal cycles. A harmonic seasonal model is first applied in BFAST to account for regular seasonal phenological variations. BFAST next computes the Ordinary Least Squares Moving Sum (OLS-MOSUM) by considering that the moving sums of the residuals after the harmonic seasonal model have been removed from the time series data values. MOSUM tests for structural change using the null hypothesis that all regression coefficients are equal i.e., every observed value can be expressed as a linear function with the same slope [25
]. If the null hypothesis is true, the values can be modeled by one line with that slope, and the sum of residuals will have a zero mean. MOSUM compares moving sums of residuals to test the likelihood of the regression coefficient for a certain time period based on a user’s input stating the minimum time between potential “breakpoints”. A rejection of the null hypothesis indicates that the regression coefficient changes at that point in time.
The MOSUM uses a default p-value of 0.05, meaning that the probability of it detecting a structural change when none has occurred is less than 5%. If MOSUM does not detect some structural change with a confidence level of 95%, it returns a “no breakpoints” result. If MOSUM detects some structural change with a confidence level of 95%, it then processes the time series through a second test, which is used to determine where the breakpoints are located in time. The output of this function is a 95% confidence interval for each breakpoint (expressed as two date numbers that define a range, before and after a breakpoint.).
For BFAST timer-series analysis, MOD13 NDVI data values (2000 to 2017) from Alaska wetland locations were subsampled to include only the growing season values during the low snow cover period of 1 May to 1 October, leaving about 10 observations per year. If a “no data” value was present in the growing season MOD13 record, then the NDVI from the previous 16-day period was substituted. Change metrics generated by BFAST from the time series analysis results included the number of breakpoints, date of each breakpoint, and the slope of the NDVI between breakpoint dates.
The principal findings of this study were that wetland cover locations across Alaska that burned at high severity subsequently recovered their green cover seasonal profiles to relatively stable pre-fire levels in less than a decade. The large wetland fires in Alaska from 2013 to 2016 showed an incomplete greenness recovery compared to earlier fires. In the years prior to 2013, the NDVI change tended to be positive at HBS wetland elevations lower than 400 m, whereas higher elevation HBS wetland locations showed much weaker greenness recovery changes by 2017. This elevation threshold of 400 m for positive post-fire NDVI recovery is not obviously related to any known topoecological changes for high-latitude wetlands, which is a new finding that merits more field research to understand this remote sensing observation. By all accounts, this is the first statewide or regional study of wetland burning and greenness recovery for Alaska that has been published, making comparisons to previous published results of a similar nature unattainable.
Nonetheless, the outcomes of recovery and regrowth pathways after high severity burning over the next few decades will be of significant consequence to the local community members in Alaska who have depended on wetland ecosystems for subsistence hunting and trapping. In and around mesic soil locations, the deep surface organic material of low bulk density in evergreen tree stands generally precludes deciduous boreal species from establishing seedlings [10
However, relatively thin post-fire organic layer depths, such as those measured by Potter [9
] in field surveys in the Tanana Area fires of 2015, may cause notable alterations in the successional outcomes of severely burned ecosystems, including a shift from the conifer-dominated thick organic layer to an increase in the dominance of deciduous or shrub species [10
]. Barrett et al. [27
] reported that the areas with less than 3 cm of surface organic layer depth after boreal forest fires will be susceptible to deciduous-dominated regeneration, whereas areas with 3–10 cm of organic layer depth will be susceptible to co-dominant regeneration by both coniferous and deciduous trees.
The aboveground biomass levels in the tundra wetlands of Alaska have been positively correlated with NDVI, and with elevated ecosystem carbon (CO2
) fluxes, including net ecosystem production and ecosystem respiration [3
]. Anywhere that most of the live vegetation is consumed by intense wildfire, nutrients can be mobilized, surface water temperature can become elevated, and soil erosion may increase [1
]. The recovery of wetland vegetation production and live green foliage cover after wildfires can help stabilize hydrological and thermal regimes, promote biodiversity, and reduce the seepage of dissolved nutrients to adjacent fluvial systems [28
In previously published studies of satellite greenness (NDVI) in Alaska using the BFAST method, Forkel et al. [29
] found the region to be of special interest for the analysis of trend change detection, because of greening NDVI trends in the tundra ecosystems of the North Slope as well as browning trends in the interior boreal forests. These authors reported that most of the breakpoints in NDVI time series coincided with large wildfire events. As in the present study of wetland NDVI trends and wildfire, BFAST methods detected stronger greening and browning trends if snow-affected values were excluded from the analysis or when only peak seasonal NDVI values were used. Breakpoints with abrupt changes, i.e., higher magnitudes, were detected more frequently than were breakpoints with gradual changes, i.e., low magnitudes.
Forkel et al. [29
] further reported that downward (browning) trends in the NDVI between 20–30 years long occurred in some of the boreal regions of central Alaska and in southwestern Alaska, usually with uncertainties of up to four years. The detection of breakpoints in the NDVI time series in 2004 agreed with the spatial distribution of Landsat-mapped large wildfires and other field-based observations. Seasonal NDVI patterns suggested that the conifer forests that burned in 2004 tended to be replaced by broad-leaved shrubs (dwarf birch and aspen) and grasses during post-fire recovery years, which resulted in a structural change in the NDVI time series.
The region-wide BFAST analysis results from the present study similarly indicated that significant structural change (in the form of breakpoints) could be detected in the 250-m NDVI time series for the vast majority of wetland locations in the major Yukon river drainages of interior Alaska that had burned at high severity since the year 2001. The lower-than-expected number of breakpoints in the MODIS NDVI time series detected by BFAST in 2015 may be explained by (1) there being fewer dates in 2016 and 2017 to compare to pre-2015 NDVI levels than for other fire years, and (2) the presence of fires prior to 2015 at the same wetland location, which would have depressed NDVI values in 2015 more than unburned locations. The predominance of positive overall slopes in the NDVI time series of wetlands with breakpoints indicated rates of vegetation recovery that are typically in the range of 40 to 200 NDVI units per year (scaled from 0–1 by a factor or 10,000).