| Description: | We constrained this study to areas that burned as stand-replacing fire
[>92.5% of basal area killed, corresponding to a Relative differenced Normalized Burn
Ratio (RdNBR) ≥ 625; Harvey 2015] in 1988 and had not burned again through 2019 (~40% of
area within the 1988 burn perimeter; Fig. A.1). We first related satellite-derived NDVI to
field-collected postfire stem density data, then used NDVI at 30 years postfire to
identify areas that were both sparse (low postfire tree density) and reduced (relative to
pre-fire biomass). Thirty-meter-resolution Landsat Top of Atmosphere (TOA) Reflectance
imagery was retrieved from Google Earth Engine (Gorelick et al. 2017) for 2013 and 2014
(Landsat 8) to create a postfire greenest pixel composite, with two years of data needed
to collect enough cloud-free pixels. These were chosen as the years nearest the year in
which postfire stem density data were collected (2012) that avoid scanner line correction
failure on the Landsat 7 satellite, the only Landsat sensor available in 2012. Although
Landsat Surface Reflectance (SR) imagery is recommended for use in most remote sensing
contexts (Young et al. 2017), Landsat TOA imagery was used because it better explained
variation in postfire stem density compared to SR imagery (Table A.1). The NDVI was
calculated by the normalized difference between bands 5 (near infrared) and 4 (red) for
all pixels within and adjacent to YNP. Because of minor differences between sensors,
Landsat 8 NDVI values were harmonized to Landsat 5 equivalents prior to analysis (Roy et
al. 2016). We removed NDVI values <0.01 and >0.9 across all images in each year to
avoid erroneous NDVI classification. We developed a postfire greenest pixel composite
using NDVI values from both “spring” (January 1 to June 5) and “fall” (September 9 to
December 31) to ensure our composite included biomass from evergreen vegetation (i.e.,
recovering coniferous forest) and excluded biomass from deciduous vegetation during
summer. Each pixel in our spring and fall composite thus represents the highest NVDI value
for that pixel across 2013-14 spring and fall dates (hereafter, evergreen NDVI). Imagery
during snow cover likely contributes little to our spring and fall composite given our
date range spans snow-free periods before green-up (late May-early June) and after
senescence (September-early October) and maximum NDVI values were used for each pixel.
Postfire tree density (stems ha-1) data were obtained from 71 plots (50 m x 50 m), all
burned as stand-replacing fire and widely distributed, that were sampled in 2012 (i.e., 24
years postfire; see Turner et al. 2016; Fig. A.2). Postfire stem density was recorded
within three belt transects (50 m x 2 m) in each plot and averaged for the plot. The
postfire spring and fall greenest pixel composite was sampled using plot centers of the 71
postfire plots, where evergreen NDVI was extracted from the 30 m x 30 m pixel that
intersected plot center. Regression analysis was used to relate 2013-14 evergreen NDVI to
postfire lodgepole pine density. Stem density was log-transformed prior to analyses to
satisfy the assumption of equal variance. Spatial structure in model residuals was
assessed with semivariograms. We also tested for relationships between pre-fire stem
density and pre-fire evergreen NDVI using a greenest pixel composite for 1986-87, but no
relationship was found (Appendix A). Distinguishing areas of poor forest recovery from
recovered forest requires determining whether postfire regeneration is both sparse and
reduced relative to pre-fire density (sensu Coop et al. 2020). Thus, postfire stem density
must be quantified throughout a burned area, and pre- and postfire densities must be
compared. We used a threshold of 1000 stems ha-1 to differentiate sparse from intermediate
and dense postfire tree density. This value is sparse relative to the mean (21,738 stems
ha-1) and median (4050 stems ha-1) stem density 24 years after the 1988 fires (Turner et
al. 2016) and is below the density to which stands converge after ~175 years of
development (1170 stems ha-1; Kashian et al. 2005). We used the regression between
postfire evergreen NDVI and postfire stem density to identify the NDVI threshold that
would distinguish present-day stands (2018-19) with sparse postfire tree densities (≤ 1000
stems ha-1) from stands where recovery was robust. To identify areas of reduced postfire
tree recovery, pre-fire (1986-87) and present-day (2018-19) spring and fall greenest pixel
composites were created using methods described above. We then subtracted pre- from
postfire evergreen NDVI and identified pixels in which postfire evergreen NDVI exhibited a
net decline >0.1 relative to pre-fire. This conservative definition of reduced recovery
targets areas with a marked decline in biomass and avoids areas with marginal changes in
evergreen NDVI that may not be ecologically meaningful. We next identified all pixels in
the present-day spring and fall composite that were (i) below the sparse recovery
threshold and (ii) had a decline in evergreen NDVI >0.1 (defined as sparse and reduced
forest recovery). We also identified all pixels that were (i) above the sparse recovery
threshold and (ii) had no decline or an increase in evergreen NDVI (defined as recovered
forest). We randomly sampled 1000 pixels within areas of both sparse and reduced recovery
and recovered forest, with all sample pixels (n = 2000) separated by ≥500 m. As indicators
of differences in vegetation conditions, we extracted the following at each sample
location: (i) net change in evergreen NDVI from pre- to postfire; (ii) vegetation height
(m), retrieved from the Global Forest Canopy Height database (Potapov et al. 2020); and
(iii) the proportion of total NDVI represented by non-evergreen vegetation [the difference
between the postfire spring and fall (2018-19) greenest pixel composite and a greenest
pixel composite at peak biomass considering all dates between January 1, 2018, and
December 31, 2019]. One hundred and forty-nine sample locations were removed from
consideration for vegetation height because data were not available. Inherent spatial and
temporal structure in remotely sensed imagery (Fig. A.3; Ives et al. 2021) and large
sample sizes pose several problems for common statistical methods, particularly in the
interpretation of statistical significance. As such, we focused our interpretation on
whether differences are ecologically meaningful rather than statistically significant. We
next extracted four potential predictors at each sample location to identify pixel- and
patch-level drivers of the distribution of sparse and reduced recovery vs. recovered
forest. Three predictors were associated with abiotic conditions that could potentially
influence tree establishment and growth [elevation (m), slope (degrees), and Topographic
Moisture Index (TMI; cosine-transformed aspect: 0 = dry, southwest facing; 2 = moist,
northeast facing; Beers et al. 1966), all derived from a 30-m resolution digital elevation
model], and the fourth predictor chosen as a proxy for seed supply (distance to ex situ
seed sources, defined as unburned areas or areas burned at low severity). Elevation is
also a proxy for pre-fire forest composition, particularly the presence of serotinous
lodgepole pines, which are common at elevations <2300 m and rare at elevations ≥2300 m
(Schoennagel et al. 2003). We did not include burn severity, because all sampled locations
burned as stand-replacing fire; pre-fire forest composition, which varies strongly with
elevation; or soil type, which does not relate to variability in postfire sapling density
(Turner et al. 2004). References Beers, T. W., P. E. Dress, and L. C. Wensel 1966. Notes
and observations: Aspect transformation in site productivity research. J For 64:691-692.
Coop, J. D., S. A. Parks, C. S. Stevens-Rumann, S. D. Crausbay, P. E. Higuera, M. D.
Hurteau, A. Tepley, E. Whitman, T. Assal, B. M. Collins, K. T. Davis, S. Dobrowski, D. A.
Falk, P. J. Fornwalt, P. Z. Fule, B. J. Harvey, V. R. Kane, C. E. Littlefield, E. Q.
Margolis, M. North, M. Parisien, S. Prichard, and K. C. Rodman. 2020. Wildfire-driven
forest conversion in western North American landscapes. BioScience 70(8):659-673.
Gorelick, N, M Hancher, M Dixon, S Ilyushchenko, D Thau, and R Moore 2017. Google Earth
Engine: Planetary-scale geospatial analysis for everyone. Remote Sens Environ 202:18-27.
Harvey, B. J. 2015. Causes and consequences of spatial patterns of fire severity in
Northern Rocky Mountain Forests: The role of disturbance interactions and changing
climate. PhD Dissertation, University of Wisconsin-Madison, Madison, Wisconsin, USA. Ives,
A. R., L. Zhu, F. Wang, J. Zhu, C. J. Morrow, and V. C. Radeloff. 2021. Statistical
inference for trends in spatiotemporal data. Remote Sens Environ 226:112678. Kashian, D.
M., M. G. Turner, W. H. Romme, and C. G. Lorimer. 2005a. Variability and convergence in
stand structural development on a fire-dominated subalpine landscape. Ecology
86(3)643-654. Potapov, P, X Li, A Hernandez-Serna, A Tyukavina, MC Hansen, A Kommareddy, A
Pickens, S Turubanova, H Tang, CE Silva, J Armston, R Dubayah, JB Blair, and M Hofton
2020. Mapping and monitoring global forest canopy height through integration of GEDI and
Landsat data. Remote Sens Environ:112165 https://doi.org/10.1016/j.rse.2020.112165. Roy,
D. P., V. Kavalskyy, H. K. Zhang, E. F. Vermote, L. Yan, S. S. Kumar, and A. Egorov. 2016.
Characterization of Landsat-7 to Landsat-8 reflective wavelength and normalized difference
vegetation index continuity. Remote Sens Environ 185:57-70. Schoennagel, T., M. G. Turner,
and W. H. Romme. 2003. The influence of fire interval and serotiny on postfire lodgepole
pine density in Yellowstone National Park. Ecology 84(11):2967-2978. Turner, M. G., D. B.
Tinker, W. H. Romme, D. M. Kashian, and C. M. Litton. 2004. Landscape patterns of sapling
density, leaf area, and aboveground net primary production of postfire lodgepole pine
forests, Yellowstone National Park (USA). Ecosystems 7:751-775. Turner, M. G., T. G.
Whitby, D. B. Tinker, and W. H. Romme. 2016. Twenty-four years after the Yellowstone
Fires: are postfire lodgepole pine stands converging in structure and function? Ecology
97(5):1260-1273. Young, NE, RS Anderson, SM Chignell, AG Vorster, R Lawrence, and PH
Evangelista 2017. A survival guide to Landsat preprocessing. Ecology
98(4):920-932. |
