Remote Sensing of Environment 121 (2012) 61–68

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r .com/ locate / rse

Analysis of sapling density regeneration in Yellowstone National Park withhyperspectral remote sensing data

Christopher Potter a,⁎, Shuang Li a,b, Shengli Huang a,c, Robert L. Crabtree c,d

a NASA Ames Research Center Mail Stop 242-4, Moffett Field, CA 94035 USAb College of Environment and Planning, Henan University, Henan 475001 Chinac Yellowstone Ecological Research Center, Bozeman, MT 59718, USAd HyPerspectives Inc., Bozeman, MT 59718, USA

⁎ Corresponding author. Tel.: +1 650 604 6164.E-mail address: Chris.Potter@nasa.gov (C. Potter).

0034-4257/$ – see front matter. Published by Elsevier Idoi:10.1016/j.rse.2012.01.019

a b s t r a c t

a r t i c l e i n f o

Article history:Received 28 June 2011Received in revised form 17 January 2012Accepted 21 January 2012Available online 17 February 2012

Keywords:Sapling densityPost-fire ecosystemsAVIRISHyperspectralFire ecologyNorthern Rocky MountainsYellowstone National Park

The density of lodgepole pine (Pinus contorta) sapling regeneration was mapped in areas burned during the1988 wildfires across Yellowstone National Park (YNP), Wyoming, USA. Hyperspectral image analysis andfield measurements were combined across the entire YNP extent. Airborne Visible Infrared Imaging Spectrometer(AVIRIS) image data from 2006 were used to compute ten different vegetation indices (VI). The ten VIs were com-bined to buildmultiple regressionmodels for predicting andmapping post-fire sapling density. Four different formsof regression modeling were applied to derive the highest possible prediction accuracy (correlation coefficient ofR2=0.83). Pine sapling regeneration 19 years after large-scale wildfires showed a high level of variability in patchdensity (ranging from 14/100 ha to 57/100 ha), whereas sapling density measured previously from the first decadefollowing wildfire was more uniform (10/100 ha to 21/100 ha). The ecosystem-level clumpiness index showedmajor shifts in aggregation of different sapling density classes, and was consistent with an overall decrease in esti-mated sapling density of nearly 50% between 1998 and 2007. This analysis revealed important succession patternsand processes in post-fire forest regeneration for the Greater Yellowstone Area (GYA).

Published by Elsevier Inc.

1. Introduction

Fire plays a major role in shaping the forested landscapes of thenorthern Rocky Mountains region (Romme and Despain, 1989). In1988, wildfires in and around Yellowstone National Park (YNP) repre-sented the largest single forest burn event in the recorded history ofthe United States up until that time (Schullery, 1989). Thirty-six per-cent of the YNP area was burned by the wildfires (Turner et al., 2003).The 1988 fires produced a 250,000-ha landscape mosaic of lodgepolepine (Pinus contorta var. latifolia Englem. ex Wats.) of different standages, with regrowing seedling densities ranging from fewer than50 stems ha−1 to more than 500,000 stems ha−1 (Turner et al.,2004). These fires represented a natural disturbance event that occursat intervals of 100–300 years in the Northern Rocky Mountains region(Romme and Despain, 1989). Smaller fires (usually less than 5000 ha)occur more frequently on the Yellowstone landscape during the inter-val between these large fires (Despain, 1990). As a result of this fireregime, forest age in YNP is highly variable, ranging from youngstands created by the 1988 fires to mature stands greater than450 years old.

nc.

During the period 1988 to 2008, sapling regeneration has struc-tured the new post-fire ecosystems within the Yellowstone land-scape, which in turn can create new wildlife habitat, restore fuelloads, and increase carbon sequestration on the land (Despain,1990; Knight, 1987; Romme, 1982; Romme and Despain, 1989).Ryan et al. (2008) reported that carbon in subalpine forests of YNP re-covers quickly (within 50 years) after a fire. Kashian et al. (2006) con-cluded that wildfire is unlikely to change carbon stored in forests bymore than 10%, unless fire permanently converts forests to grass-lands. Potter et al. (2011) modeled ecosystem production and carbonfluxes in YNP and showed that most of the areas burned in 1988 willcontinue to be a net carbon flux to the atmosphere for decades tocome. They also suggested that airborne and satellite observationsat high spatial resolutions (b100 m) for vegetation structure andwildfire recovery patterns would be the next necessary researchstep toward carbon budget refinement, due to limited ground-basedmeasurement data sets over vast, remote areas of YNP.

Regeneration mechanisms of tree saplings in post-fire ecosystemshave been well-documented (Cierjacks et al., 2008; Costas et al.,1996; Spanos et al., 2000; Kruger and Reich, 1997; Mlambo andMapaure, 2006; Thanos and Marcou, 1991; Turner et al., 1997;Vandvik et al., 2005). Numerous abiotic (e.g., elevation, aspect,slope, temperature, soil type, localized precipitation, burn severity)and biotic factors (e.g., plant species and pre-fire stand age) interact

Fig. 1. Study area with burned areas of YNP from the 1988 wildfires. Burn severity was derived from the Landsat Relative differenced Normalized Burn Ratio (RdNBR, Key and Benson,2005; Miller et al., 2009). Imagery dates used for this RdNBR product were pre-fire: July 22, 1988 and post-fire: October 10, 1988.

62 C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

to determine the patterns of sapling density. Lodgepole pine, whichdominates the forest cover in YNP, is a serotinous evergreen treethat produces closed cones and releases its seeds only when heated byfires. Saplings of lodgepole pine often regenerate in dense, even-agedstands (Turner et al., 2004).

Previous field research in YNP has focused on specific tree species inearly (2–10 years) post-fire regeneration phases (Turner et al., 1997),especially those in nutrient-poor ryholitic soils in the Yellowstonecaldera. These early studies used ground truth measurements orstandardized transect-based methods to investigate the re-establishment of post-fire forest stands and infer regenerationmechanisms from both biotic and abiotic factors.

Aerial photography and color infrared orthophotographs wereused by Turner et al. (2004) and Kashian et al. (2004) to extract sap-ling density distribution patterns, aboveground net primary produc-tion, and LAI (Leaf Area Index) over relatively small (0.25 ha) plotareas 10 years after the 1988 fires. By 1999, estimated sapling densi-ties ranged from 0 to >500,000 stems ha−1, although >60% of theburned area was represented by densities b5000 stems ha−1. Burnedforests areas consisted of small regenerating patches averaging 1.5 ha.High-density sapling patcheswere spatially coincident with severe sur-face fire areas, suggesting that burn severity, stand age, and serotinymay be important and interacting determinants of spatial variation insapling density.

The goal of our study was to better understand the spatial patternsof sapling density regeneration in the second decade after the 1988fires throughout the extent of YNP. Ground-truth measurements ina total of 192 burned forest plots were made for the purpose of pre-dicting and validating remote sensing estimates of sapling density.Due to spectral limitations of satellite imagery (Landsat and SPOT;Shaw et al., 1998), AVIRIS imagery was collected in 2006 and ana-lyzed. Ten different AVIRIS vegetation indices (VIs) were combinedto build multiple regression models for mapping post-fire saplingdensity. Several statistical indices of spatial pattern were computedfor further comparison of the 1998 and 2007 estimates of sapling re-generation in YNP.

2. Study area

Our study area covered the entire extent of YNP and its immediatelysurrounding lands in northwest Wyoming, southwest Montana, andsoutheast Idaho, USA (Fig. 1). The YNP area has elevations rangingfrom 1540 m to 3760 m. Nearby mountain ranges include the GallatinRange to the northwest, the Beartooth Mountains in the north, theAbsaroka Range to the east, and the Teton Range and the MadisonRange to the southwest and west. The Continental Divide of NorthAmerica runs diagonally through the southwestern part of YNP. The di-vide is a topographic feature that separates Pacific Ocean and AtlanticOcean water drainages. About one third of YNP lies on the west sideof the divide. The climate is generally cool and dry with mean Januaryand July temperatures of −11.4 °C and 10.8 °C, respectively, andmean annual precipitation of 56.3 cm (Dirks and Martner, 1982). Win-ters are long and cold, lasting frommid-November tomid-March. Sum-mers are short and often dry, usually lasting from July through August.Average annual snow depth is around 33 cm (Huang et al., 2008).

Soils in YNP at higher elevations are largely nutrient-poor rhyo-lites and andesites with low water-holding capacity (Rodman et al.,1996). Valley bottoms and floodplains contain glacial out-wash andalluvial soils that are higher in nutrients and water-holding capacity.Soils derived from rhyolitic parent materials typically are coarser andhave fewer base cations and lower water-holding capacity than soilsderived from andesite or lacustrine sediments. Lacustrine sedimentstypically have the highest silt and clay content, base cations, andwater-holding capacity (Despain, 1990).

The forests of YNP consist of five main conifer species (Kokalyet al., 2003): lodgepole pine, whitebark pine (Pinus albicaulis), Douglasfir (Pseudotsuga menziesii), Engelmann spruce (Picea engelmannii), andsubalpine fir (Abies lasiocarpa). Elevation and soil fertility are consideredto be the twomost important abiotic gradients controlling forest vegeta-tion on the subalpine plateaus (Crabtree et al., 2009; Despain, 1990).Non-forest vegetation is divided into fourmajor groups: grassland, sage-brush steppe (shrubland), wet sedge and willow meadow, and alpinemeadow. The distributions of these vegetation types are strongly

Table 1AVIRIS vegetation index (VI) definitions.

VIAbbreviation

Definition

NDVI [NIR−Red]/[NIR+Red]SR NIR/RedEVI 2.5 [NIR−Red]/[NIR+6 Red−7.5 Blue]ARVI [NIR−(2 Red−Blue)]/[NIR+(2 Red−Blue)]NDVI705 [750 nm−705 nm]/[750 nm+705 nm]mSR705 [750 nm−445 nm]/[705 nm−445 nm]mNDVI705 [750 nm−705 nm]/[750 nm+705 nm−2 445 nm]VOG1 740 nm/720 nmNDNI [log(1/1510 nm)− log (1/1680 nm)]/[log(1/1510 nm)+log

(1/1680 nm)]MSI 1599 nm/819 nm

NIR band reflectance is centered at 800 nm.Red band reflectance is centered at 680 nm.Blue band reflectance is centered at 450 nm.

63C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

influenced by elevation. Big sagebrush grows in dry to moderately moistareas atmiddle and lower altitudes, such as in the Lamar River Valley. Sil-ver sage grows in wetter areas higher than 2130 m, for example, in theHayden and Pelican Valleys. Alpine meadows are present at elevationshigher than 3050 m. Sedgemarshes and other wetland vegetation thrivein areas of year-round standing water at various elevations throughoutYNP (Kokaly et al., 2003).

Most of YNP is presently managed as a wilderness area, wherepost-fire patterns have been little affected by direct human actions(Turner et al., 2004). The 1988 fires burned over 2500 km2 in YNPand surrounding lands (Fig. 1). The summer of 1988 was the driestin more than a century, with high temperatures and extremely highwinds produced by a series of dry high-pressure systems (Baskin,1999). Wildfires in 1988 spread rapidly through all forest successionalstages and appeared to be influencedmore bywind speed and directionthan by subtle distribution patterns in fire fuels or topography (Turneret al., 1997). This created amosaic of burn severities, including light sur-face burn, severe surface burn, and crown fire, wherein the latter twoclasses the canopy trees were killed (Christensen et al., 1989; Turneret al., 1997).

3. Methods

3.1. Field surveys

In the July-September period of 2007, 186 17-m radial plots wereestablished in areas impacted by the 1988 wildfires. Within plots,center locations were recorded using post-differentially correctedTrimble GPS measurements, accurate to within ±2 m. (Huang et al.,2009). In each plot, pine sapling numbers were counted manually.Small saplings with a height b25 cm were not recorded in the counts.At the center of each plot, digital photographs were taken in each car-dinal direction to later verify sapling numbers measured in the field.In 76 plots where the sapling density was too high to permit exten-sive movement between trees stems, we counted the sapling numberwithin a circular plot radius of 5 m, and normalized the count numberto a 17-m plot radius. An additional six plots were surveyed in 2008,resulting in a total of 192 plots used in this study.

3.2. Hyperspectral datasets

AVIRIS image data were collected by NASA/JPL on September 25,2006 in 224 contiguous channels of approximately 10-nm bandwidthover the spectral wavelength range of 350–2500 nm. The spatial resolu-tion on the ground was around 17-m at nadir. The orthocorrection sys-tem for AVIRIS processing employed full three-dimensional ray tracingand a 30-m spatial resolution digital elevation model for completeorthorectification. Each pixel in the image was individually ray traced

using the best estimate of sensor location and attitude until it inter-sected the DEM. The spatial fidelity of the data was much improvedfromprevious AVIRIS datasets, especially in areas of rugged and variableacross-track terrain, resulting in an accuracy of one pixel (Schläpferet al., 2000). The uncorrected AVIRIS radiance datawere atmosphericallycorrected and converted to reflectance in ACORN software (Huang et al.,2009, 2010).

3.3. AVIRIS Vegetation Index (VI) descriptions

More than 150 VIs have been published in scientific literature, butonly a subset have been systematically tested for use in predictingsapling densities of regenerating forest stands. Our selection of themost appropriate indices for our study was based upon robustness,scientific basis, and general applicability for detection of the presenceand relative abundance of pigments, water, and carbon as expressedin the solar-reflected optical spectrum of 400 nm to 2500 nm (Asneret al., 2000). Moreover, our AVIRIS imagery was acquired on Septem-ber 25, 2006. During this period, pine saplings were still green whilethe understory herbaceous vegetation would be senesced to brown.

Ten of the most commonly used VI formulations (previouslyreviewed by Ceccato et al., 2001; Elvidge and Chen, 1995; Hueteet al., 2002; Kaufman and Tanré, 1992; Serrano et al., 2002; Tucker,1979, and Vogelmann et al., 1993) that would enhance the contrastbetween live-green sapling cover and dead-brown grass were calcu-lated from AVIRIS bands (Table 1, with wavelength numbers in nm)and can be described as follows.

VI1—Normalized Difference Vegetation Index (NDVI) is one ofmost well-known, and most frequently used VIs computed frommulti-spectral remote sensing (Tucker, 1979). The normalized combi-nation of the near infrared (NIR) and red wavelength bands capturesthe reflectance zones for chlorophyll and makes NDVI robust over awide range of vegetation cover conditions. NDVI can, however, satu-rate in dense vegetation cover conditions when leaf area (LAI) be-comes high.

VI2—Simple Ratio (SR) index is another well-known multi-spectral VI (Tucker, 1979). SR is the ratio of NIR and red reflectancebands and, as with the NDVI, SR can saturate in dense vegetationwhen LAI becomes high.

VI3—Enhanced Vegetation Index (EVI) was developed to optimizethe vegetation signal in high LAI cover by using the blue reflectancebands with the near infrared (NIR) and red wavelength bands to cor-rect for soil background signals and reduce atmospheric interference,including aerosol scattering (Huete et al., 2002).

VI4—Atmospherically Resistant Vegetation Index (ARVI) is an en-hancement to the NDVI that is relatively resistant to atmospheric fac-tors (for example, aerosol). It uses the reflectance in the blue bands tocorrect the red reflectance for atmospheric scattering (Kaufman andTanré, 1992). ARVI is most useful in regions of high atmospheric aero-sol content, including tropical regions contaminated by soot fromslash-and-burn agriculture.

VI5—Red Edge Normalized Difference Vegetation Index (NDVI705)is a modification of the traditional NDVI intended for use with highspectral resolution reflectance data, such as AVIRIS. This VI differsfrom the NDVI by including bands along the red edge, instead of themain red absorption and reflectance peaks (Sims and Gamon, 2002).The NDVI705 capitalizes on the sensitivity of the vegetation rededge to small changes in canopy foliage content, gap fraction, andsenescence.

VI6—Modified Red Edge Simple Ratio Index (mSR705) is a modifi-cation of the traditional SR index that uses bands along the red edgeto incorporate corrections for leaf specular reflection (Sims andGamon, 2002).

VI7—Modified Red Edge Normalized Difference Vegetation Index(mNDVI705) is a modification of the NDVI705 that incorporates a cor-rection for leaf specular reflection (Sims and Gamon, 2002).

Table 2Step-wise linear regressionmodel results for prediction of 2007measured sapling density(SD) in YNP to AVIRIS VIs. All VI coefficient values listed were entered with a confidencelevel of pb0.05.

Modelform

Best-fit equation R2

SD SD=138,656+20,099∗SR−217,848∗VOG1+395,190∗NDVI705+235,205∗NDNI−126,225∗ARVI+176,085∗EVI

0.78

Ln(SD) Ln(SD)=21.923+62.977∗NDVI705−22.221∗VOG1 0.80Ln(SD)−Ln(VI)

Ln(SD)=19.726+19.231∗Ln(NDVI705)+9.862∗Ln(SR)−28.927∗Ln(VOG1)−17.019∗Ln(NDVI)

0.83

Sqrt(SD) Sqrt(SD)−576.23+60.25∗SR−888.96∗VOG1+1923.90∗NDVI705+811.91∗NDNI−410.49∗ARVI+613.25∗EVI

0.79

20 55 150 400 1100 3000 8100 22000 60000 163000

Fig. 2. Comparison of predicted SD values fromAVIRIS VIs (natural log transform regression)to field-measured SD counts in YNP from 2007 to 2008. Approximate SD values have beenconverted to original units of stems ha−1 below the x-axis.

Table 3Best-fit prediction of SD values from AVIRIS VI regression distributed according to twodifferent classification schemes.

Class Turner et al. (2004)classification (stems ha-1)

Percent Geometrical intervalmethod (stems ha−1)

Percent

1 b1000 54.4 b1507 59.42 1000–5000 21.0 1507–1756 2.03 5000–10,000 7.4 1756–3263 8.64 10,000–15,000 3.5 3263–12,371 14.75 15,000–50,000 7.3 12,371–67,409 10.16 >50,000 6.4 67,409–400,000 5.2

64 C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

VI8—Vogelmann Red Edge Index 1 (VOG1) is a narrow-band re-flectance measurement that is sensitive to the combined effects of fo-liage chlorophyll concentration, canopy leaf area, and water content(Vogelmann et al., 1993).

VI9—Normalized Difference Nitrogen Index (NDNI) is designed toestimate the relative amounts of nitrogen contained in chaparralshrub canopies. Together, leaf nitrogen concentration and canopy fo-liar biomass can be combined in the 1510 nm range to predict totalcanopy nitrogen content (Serrano et al., 2002).

VI10—Moisture Stress Index (MSI) is a reflectance measurementthat is sensitive to changing leaf water content. As the water contentof leaves in vegetation canopies increases, the strength of the absorp-tion around 1599 nm increases. Absorption at 819 nm is nearly unaf-fected by changing water content, so this band is used as thereference (Ceccato et al., 2001). The MSI is inverted relative to theother water VIs; higher values indicate greater water stress.

3.4. Regression approaches

All ten AVIRIS VIswere considered as candidate explanatory variablesin stepwise regression analysis that was repeated using four differentmodel forms for prediction of the measured sapling density (SD) dataset. Half of the fieldmeasurements (at 96 plots) were randomly selectedto calibrate the regression model results.

First, multiple regression with a simple linear model was applied,denoted as:

SD ¼ β0þ β1 � VI1þ β2 � VI2þ β3 � VI3þ β4VI � 4þ β5 � VI5þβ6 � VI6þ β7 � VI7þ β8 � VI8þ β9 � VI9þ β10 � VI10þ ε

Second, a multiplicativemodel formwas applied. The sapling densitywas natural log-transformed and estimated as a linear model of ten VIsvariables

Ln SDð Þ ¼ β0þ β1 � VI1þ β2 � VI2þ β3 � VI3þ β4 � VI4þ β5 � VI5þβ6 � VI6þ β7 � VI7þ β8 � VI8þ β9 � VI9þ β10 � VI10þ ε

Third, a multiplicative model with the sapling density being naturallog-transformed was estimated as a linear model, denoted as:

Ln SDð Þ ¼ Ln β0ð Þ þ β1 � Ln VI1ð Þ þ β2 � Ln VI2ð Þ þ β3 � Ln VI3ð Þ þβ4 � Ln VI4ð Þ þ β5 � Ln VI5ð Þ þ β6 � Ln VI6ð Þ þ β7 � Ln VI7ð Þ þ β8 �Ln VI8ð Þ þ β9 � Ln VI9ð Þ þ β10 � Ln VI10ð Þ þ ε

Fourth, a multiplicative model with the sapling density in squareroot transformation was estimated as a linear model, denoted as:

Sqrt SDð Þ ¼ β0þ β1 � VI1þ β2 � VI2þ β3 � VI3þ β4 � VI4þ β5 �VI5þ β6 � VI6þ β7 � VI7þ β8 � VI8þ β9 � VI9þ β10 � VI10þ ε

Spatial patterns in the resulting SD products were computed usingFRAGSTATS (McGarigal and Marks, 1995) and the Geometrical Inter-val Method (GIM; Wu and Rao, 2006). The GIM was adopted to splitthe data set into six classes of SD values. GIM is a classificationscheme wherein the class breaks are based on a geometrical seriesthat depends upon a constant coefficient multiplier for each value inthe series. Thismethod is useful for classifying non-normally distributedoutput data and ensures that each class range has approximately thesame number of values within each class.

4. Results and discussion

Sapling density measured in field survey plots ranged from a maxi-mum of 57,473 ha−1 to a minimum of 33 ha−1, with a median value of1604 ha−1. Step-wise linear regression modeling for prediction of SDresulted in significant correlation coefficient values (R2>0.78;pb0.001) using any of the four different model forms (Table 2). Whilethe multiplicative model with SD being natural log-transformedshowed the highest correlation coefficient value of R2>0.83 of thefour model forms, the most significant VI predictor variables selectedby the step-wise procedure were consistent in all the model forms,

Fig. 3. Predicted sapling density map (with Yellowstone National Park boundary in blue), rendered by the Geometrical Interval Method from the natural log transform stepwiseregression model VI form shown in Table 2. Flight line gaps in the AVIRIS data collection were set to missing SD values.

65C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

primarily the VIs computed as SR, VOG1, NDVI1705, and secondarily,the VIs computed as EVI, ARVI, NDNI, and NDVI. These overall resultsimply that information about the red-edge reflectance, together withfoliage chlorophyll concentration, canopy leaf area, and water content

0

100000

200000

300000

400000

Elevation

Sap

ling

Den

sity

ha-

1

0

1000

2000

3000

4000

5000

3500

Elevation

Sap

ling

Den

sity

ha-

1

1500 2000 2500 3000 3500

1500 2000 2500 3000

Fig. 4. Distribution of predicted SD from AVIRIS VI best-fi

provided by high-spectral resolution AVIRIS imagery, were major con-tributors to the best regression modeling results for SD predictions.

The remaining half of the field measurements (for 96 plots) wasused for validation comparison of predicted SD values using the

0

50000

100000

3500

Elevation

Sap

ling

Den

sity

ha-

1

0

500

1000

1500 2000 2500 3000

1500 2000 2500 3000 3500

Elevation

Sap

ling

Den

sity

ha-

1

t regression with elevation in four main SD classes.

Table 4Summary statistical comparison of lodgepole pine SD in YNP at 10 years and 19 yearsafter the 1988 wildfires.

Statisticfor SD

Turner et al. (2004)10 years post-burn (stems ha−1)

This study from AVIRIS 19 yearspost-burn (stems ha−1)

Maximum 535,000 400,000Mean 29,381 15,128Median 3100 659

66 C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

Ln(SD)–Ln(VI) transform model form. This comparison revealed aclose match along the 1:1 line (Fig. 2). The distribution of residualerror for this final model result was very close to normal. The best-fit prediction using selected AVIRIS VIs to SD plot values was signifi-cant at R2>0.88 (pb0.01). For prediction of low SD levels(b1000 stems ha−1), the absolute error result of 3930 stems ha−1

limits the regression modeling accuracy. Overall, the regression pre-diction method showed a minor tendency to overestimate in the pre-dicted SD values for measured plot values of lower than 1000 stemsha−1 and to underestimate in the predicted SD values for measuredplot values of lower than 8000 stems ha−1.

For further analysis over the entire study region, the GIM wasadopted to split the data set into six classes of SD values. The GIMclassification scheme for our predicted SD values has been comparedin Table 3 to the scheme used by Turner et al. (2004) for measure-ments of SD in YNP, showing that the GIM classification scheme bet-ter characterized the full frequency distribution of SD estimates inYNP.

The best-fit prediction of SD values derived from the AVIRIS VIsgave a range from zero to 400,000 stems ha−1, with a predictedmean value of 15,130 stems ha−1 and standard deviation of 54,570.Approximately 96% of the area burned by 1988 wildfires was coveredin 2006 by a predicted SD of lower than 100,000 stems ha−1 (Fig. 3),while slightly more than 1% of the area burned in 1988 had a pre-dicted SD greater than 400,000 stems ha−1

Turner et al. (2004) reported that variation in the spatial distribu-tion of regenerating lodgepole pine SD measured at 10 years after the1988 wildfires was explained largely by elevation differences in theplot locations sampled, with SD decreasing significantly with increas-ing elevation starting above 2000 m. For comparison at 19 years afterthe 1988 wildfires, we selected 2000 random points from each ofmain four SD classes to examine the predicted SD variation with ele-vation (Fig. 4). In our 2006 results, we could detect no significant

Fig. 5. Spatial distribution patterns of SD estimated by (A) Turner et al. (2004)

linear correlation with elevation in any of the four main SD classes.We hypothesize that forest succession had changed SD patterns at19 years post-burn, such that elevation effects were suppressed byother abiotic (soil type) or biotic (competition and plant production)factors (Crabtree et al., 2009).

Comparison between SD statistics reported by Turner et al. (2004)and our results revealed that, nearly 20 years post-burn, the meanand median SD values in YNP decreased notably (Table 4). Specifical-ly, mean SD estimates declined by nearly 50% over the period of 1998to 2007. Young lodgepole pines establish very rapidly and become re-productively active within five years (Critchfield, 1980). By around20 years of regrowth from fire, the stand canopy can close and under-story productivity can become negligible (Basile, 1975). If regeneratingpine stands in YNP did indeed reach canopy closure stage at around2007, then thinning of sapling numbers would have been expected inthe period since 1999.

By examining maps of estimated SD from 1998 color infrared aerialphotographs (Turner et al., 2004) and from 2007 (this study, Fig. 5), weidentified a number of similarities and differences in apparent SD re-generation patterns. In both maps, many of highest SD values(>10,000 stems ha−1) were identified in the North Fork Fire areas innorthwestern sections YNP forests that burned in 1988 (Fig. 1). Basedon the more recent 2007 AVIRIS image analysis, we identified manymore areas with SD >10,000 stems ha−1 in the Fan, Clover-Mist, Red-Shoshone, and Huck Fire areas than were detected in 1998 mapping. Inmost other burned areas in YNP, we detected lower SD than estimatedin 1998 (Turner et al., 2004).

Several indices of spatial pattern were computed using FRAG-STATS (McGarigal and Marks, 1995) for further comparison of the1998 and 2007 estimates of SD regeneration in YNP (Table 5). Basedon the more recent 2007 AVIRIS image analysis, patch density rangedmore widely (from 14/100 ha to 68/100 ha), compared to SD in thefirst decade following wildfire with patch density more narrowly es-timated at 10/100 ha to 21/100 ha. The mean patch size and theClumpiness index both increased notably from 1998 to 2007 in al-most every SD class, indicating a more complete mapping of regener-ating forest stands using the AVIRIS imagery, and that patchesbecame more aggregated over the period. The Clumpiness indexshows the frequency with which different pairs of patch types (in-cluding like adjacencies between the same patch type) appear side-by-side on the map. Clumpiness equals −1 when the focal patchtype is maximally disaggregated; Clumpiness equals 0 when thefocal patch type is distributed randomly, and Clumpiness approaches1 when the patch type is maximally aggregated.

and (B) AVIRIS remote sensing (this study) in 1998 and 2007, respectively.

Table 5Comparison of spatial index patterns in two stages of post-fire ecosystem succession inYNP.

Class Saplingdensity(stems ha−1)

Proportionoflandscape(% of total)

Patchdensity(No./100 ha)

Area-weightedmean patch size(ha)

ClumpinessIndex

1998 2007 1998 2007 1998 2007 1998 2007

1 b1000 36 54 18 18 179 13,423 0.30 0.802 1000–5000 30 21 21 48 51 218 0.22 0.563 5000–10,000 10 7 17 68 1.6 1.8 0.13 0.384 10,000–15,000 8 4 16 58 3.3 0.2 0.15 0.295 15,000–50,000 9 7 17 35 1.5 11.6 0.12 0.506 >50,000 7 6 10 14 5.9 90 0.25 0.73

Spatial pattern analyses were conducted using FRAGSTATS (McGargigal and Marks 1995)with a resolution of 14.3 m×14.3 m (1998 data from Turner et al., 2004).

67C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

To better understand the effects of forest stand history on SD re-generation patterns in YNP in the second decade following the 1988wildfires, we overlaid the AVIRIS-regression SD data in Fig. 5b withthe YNP forest class map of Despain (1990). Renkin and Despain(1992) provided a definition of YNP forests stands of lodgepole pine(lp) as pioneer, early, middle, and late (lp0, lp1, lp2, and lp3, respec-tively) successional stages according to age (0–50, 50–150, 150–300,and >300 years, respectively). Results showed that early and middlesuccessional classes (50–300 years old) had a significantly higher(pb0.01) mean SD, with over 50,000 stems ha−1, compared to anyof the other (pioneer or late successional) pre-1988 forest classes(Fig. 6). All pre-1988 late successional forest classes had mean SDlower than 10,000 stems ha−1.

5. Conclusions

The 1988 Yellowstone fires produced a complex spatial mosaic ofburn severity classes on the landscape (Johnson and Miyanishi, 2007;Schullery, 1989), many of which were found to be regenerating withan estimated sapling density still ranging from less than 1000 stemsha−1 up to 400,000 stems ha−1. This study demonstrated that air-borne hyperspectral imagery collected in late summer can detectthe wide variations in measured sapling density with a high level ofconfidence. Patch sizes of regenerating lodgepole pine stands haveevidently become larger and more aggregated over the past decade

Fig. 6. Mean SD values computed from the data values in Fig. 5b for pre-1988 forestclasses specified by Despain (1990) in YNP. Two standard errors of the mean areshown in the bars for each forest class. Key to forest classes from Despain (1990) andRenkin and Despain (1992). lp0, post-disturbance lodgepole pine; lp0/nf, post-distur-bance lodgepole pine and non-forest; lp1, early successional lodgepole pine; lp1/df,early successional lodgepole pine and climax Douglas-fir; lp1/lp2, early and middlesuccessional lodgepole pine; lp1/nf, early successional lodgepole pine and non-forest;lp2, middle successional lodgepole pine; lp2/lp3, middle and late successional lodge-pole pine; lp2/nf, middle successional lodgepole pine and non-forest; lp/lp2, middleand climax lodgepole pine; lp, climax lodgepole pine; lp/lp3, climax and late succes-sional lodgepole pine; lp1/lp, climax and post-disturbance lodgepole pine; lp3, latesuccessional lodgepole pine; lp3/nf, late successional lodgepole pine and non-forest;lpp, pygmy lodgepole pine; lpp/lp3, pygmy and late successional lodgepole pine.

(1998–2007) of recovery from the 1988 fires. The rate and patternof forest regeneration from wildfires in YNP will continue to be afunction of the history of fire size, severity, return interval, and char-acteristics of the previous forest stand (e.g., cone serotiny, tree ageand stem density, and dead wood biomass). Serotiny is the long-term storage of seeds within mature closed cones which, when heat-ed by fire, open to release seeds onto newly exposed mineral soil.Lodgepole pine stands predominantly exhibit either open or closedcones and the portion of trees in a stand area with the ability to pro-duce closed cones during their lifetime as the “percent serotiny”. Thecomprehensive mapping of YNP reported in this study should permitmore in-depth testing of the hypothesis that the highest SD regener-ation rates will occur in 50–300 year old pre-fire lodgepole pinestands where percent serotiny is high and fire severity may be mod-erate, whereas the lowest SD regeneration rates will occur in areasof pre-fire late successional (>300 year old) forests, where crownfires may be common and pre-fire serotiny is low.

Acknowledgments

The authors thank themany field crews, especially Jamie Robertson,Nate Emery, Amelia Hagen-Dillon, and Jeanine Moy, of the YellowstoneEcological Research Center for their dedicated field survey work. Wethank Alan Swanson on his assistance on statistical analysis. We aregrateful to Ann Rodman, PJ White, and Roy Renkin of the YellowstoneCenter for Resources, National Park Service, for their guidance on datausage and research planning. This research was supported by appoint-ments to the NASA Postdoctoral Program at the NASA Ames ResearchCenter (for both authors Shuang Li and Shengli Huang) administeredby Oak Ridge Associated Universities through a contract with NASA.

References

Asner, G. P., Wessman, C. A., Bateson, C. A., & Privette, J. L. (2000). Impact of tissue, canopy,and landscape factors on the hyperspectral reflectance variability of arid ecosystems.Remote Sensing of Environment, 74(1), 69–84.

Basile, J. V. (1975). Forage productivity in the lodgepole pine type. InDavidM. Baumgartner(Ed.),Management of lodgepole pine ecosystems: Proceedings of a symposium; 1973October 9–11; Pullman,WA (pp. 246–263). Pullman,WA:Washington State University.

Baskin, Y. (1999). Yellowstone fires: A decade later. BioScience, 49(2), 93–97.Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., & Gregoire, J. M. (2001). Detecting

vegetation leaf water content using reflectance in the optical domain. Remote Sensingof Environment, 77, 22–33.

Christensen, N. L., Agee, J. K., Brussard, P. F., Hughes, J., Knight, D. H., Minshall, G. W.,Peek, J. M., Pyne, S. J., Swanson, F. J., Thomas, J. W., Wells, S., Williams, S. E., & Wright,H. A. (1989). Interpreting the Yellowstone fires of 1988. BioScience, 39, 678–685.

Cierjacks, A., Ruhr, N. K., Wesche, K., & Hensen, I. (2008). Effects of altitude and livestock onthe regeneration of two tree line forming Polylepis species in Ecuador. Plant Ecology,194, 207–221.

Costas, A. T., Daskalakou, E. N., & Nikolaidou, S. (1996). Early post-fire regeneration of aPinus halepensis forest on Mount Párnis, Greece. Journal of Vegetation Science, 7(2),273–280.

Crabtree, R., Potter, C., Mullen, R., Sheldon, J., Huang, S., Harmsen, J., Rodman, A., & Jean, C.(2009). A modeling and spatio-temporal analysis framework for monitoring environ-mental change using NPP as an ecosystem indicator. Remote Sensing of Environment,113, 1486–1496.

Critchfield, W. B. (1980). Genetics of lodgepole pine. U.S. Dep. Agric., For. Serv., Res. Pap.WC-37.

Despain, D. G. (1990). Yellowstone's vegetation: The consequences of history and environmentin a natural setting. New York: Robbers Rinehart, Inc.

Dirks, R. A., & Martner, B. E. (1982). The climate of Yellowstone and Grand Teton NationalParks. Occasional Paper No. 6, U. S. National Park Service, Washington, D.C.

Elvidge, C. D., & Chen, Z. (1995). Comparison of broad-band and narrow-band red andnear-infrared vegetation indices. Remote Sensing of Environment, 54, 38–48.

Huang, S., Crabtree, R. L., Potter, C., & Gross, P. (2009). Estimating the quantity andquality of coarsewoodydebris inYellowstone's post-fire forest ecosystem from fusionof SAR and optical data. Remote Sensing of Environment, 113, 1926–1938.

Huang, S., Potter, C., Crabtree, R. L., Hager, S., & Gross, P. (2010). Fusing optical andradar data to estimate sagebrush, herbaceous, and bare ground cover in Yellowstone.Remote Sensing of Environment, 114(2), 251–264.

Huang, S., Rich, P., Crabtree, R., Potter, C., & Fu, P. (2008). Modelling near-surface airtemperature from solar radiation and lapse rate: Application over complex terrainin Yellowstone National Park, USA. Physical Geography, 29(2), 158–178.

Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002). Overviewof the radiometric and biophysical performance of the MODIS vegetation indices.Remote Sensing of Environment, 83, 195–213.

68 C. Potter et al. / Remote Sensing of Environment 121 (2012) 61–68

Johnson, E. A., & Miyanishi, K. (Eds.). (2007). Plant disturbance ecology: The process andthe response. San Diego: Academic Press.

Kashian, D. M., Romme, W. H., Tinker, D. B., Turner, M. G., & Ryan, M. G. (2006). Carbonstorage on coniferous landscapes with stand-replacing fires. BioScience, 7,598–606.

Kashian, D. M., Tinker, D. B., Turner, M. G., & Scarpace, F. L. (2004). Spatial heterogeneityof lodgepole pine sapling densities following the 1988 fires in Yellowstone NationalPark, Wyoming, USA. Canadian Journal of Forest Research, 34, 2263–2276.

Kaufman, Y. J., & Tanré, D. (1992). Atmospherically resistant vegetation index (ARVI)for EOS-MODIS. IEEE Transactions on Geoscience and Remote Sensing, 30, 261–270.

Key, C. H., & Benson, N. C. (2005). Landscape assessment: ground measure of severity,the Composite Burn Index; and remote sensing of severity, the Normalized BurnRatio. In D. C. Lutes, R. E. Keane, J. F. Caratti, C. H. Key, N. C. Benson, & L. J. Gangi(Eds.), FIREMON: Fire Effects Monitoring and Inventory System USDA Forest Service,Rocky Mountains Research Station General Technical Report.

Knight, D. H. (1987). Parasites, lightning, and the vegetation mosaic in wilderness land-scapes. In M. G. Turner (Ed.), Landscape heterogeneity and disturbance (pp. 59–84).New York, New York: Springer-Verlag.

Kokaly, R. F., Despain, D. G., Clark, R. N., & Livo, K. E. (2003). Mapping vegetation inYellowstone National Park using spectral feature analysis of AVIRIS data. RemoteSensing of Environment, 84, 437–456.

Kruger, E. L., & Reich, P. B. (1997). Responses of hardwood regeneration to fire in mesicforest openings. Canadian Journal of Forest Research, 27(11), 1822–1831.

McGarigal, K., & Marks, B. J. (1995). FRAGSTATS. Spatial analysis program for quantifyinglandscape structure. USDA Forest Service General Technical Report PNW-GTR-351.

Miller, J. D., Knapp, E. E., Key, C. H., Skinner, C. N., Isbell, C. J., Creasy, R. M., & Sherlock, J.W.(2009). Calibration and validation of the relative differenced Normalized BurnRatio (RdNBR) to three measures of fire severity in the Sierra Nevada and KlamathMountains, California, USA. Remote Sensing of Environment, 113(3), 645–656.

Mlambo, D., & Mapaure, I. (2006). Post-fire resprouting of Colophospermum mopanesaplings in a southern African savanna. Journal of Tropical Ecology, 22(2), 231–234.

Potter, C., Klooster, S., Crabtree, R., Huang, S., Gross, P., & Genovese, V. (2011). Carbonfluxes in ecosystems of Yellowstone National Park predicted from remote sensingdata and simulation modeling. Carbon Balance and Management, 6, 3, doi:10.1186/1750-0680-6-3.

Renkin, R. A., & Despain, D. G. (1992). Fuel moisture, forest type, and lightning-causedfire in Yellowstone National Park. Canadian Journal of Forest Research, 22, 37–45.

Rodman, A., Shovic, H. F., & Thomas, D. (1996). Soils of Yellowstone National Park.Yellowstone National Park, Yellowstone Center for Resources. Mammoth, WY324 pp.

Romme, W. H. (1982). Fire and landscape diversity in subalpine forests of YellowstoneNational Park. Ecological Monographs, 52, 199–221.

Romme, W. H., & Despain, D. G. (1989). Historical perspective on the Yellowstone Firesof 1988. BioScience, 39(10), 696–699.

Ryan, M. G., Kashian, D. M., Romme, W. H., Turner, M. G., Smithwick, E. A. H., & Tinker,D. B. (2008, Sept. 22–27). Fire and carbon cycling for theYellowstoneNational Park land-scape. Proceedings of the IAWFMeeting, The '88 Fires: Yellowstone and Beyond, Jackson,WY.

Schläpfer, D., Hausold, A., & Richter, R. (2000). A unified approach to parametric geocodingandatmospheric/topographic correction for wide FOV Airborne Imagery. Part 1: Para-metric ortho-rectification process. Proc. 2nd EARSeL Workshop on Imaging Spectroscopy,EARSeL, Enschede (pp. 9).

Schullery, P. (1989). Yellowstone fires: A preliminary report. Northwest Science, 63(1),44–54.

Serrano, L., Penuelas, J., & Ustin, S. L. (2002). Remote sensing of nitrogen and lignin inMediterranean vegetation from AVIRIS Data: Decomposing biochemical fromstructural signals. Remote Sensing of Environment, 81, 355–364.

Shaw, D. T., Malthus, T. J., & Kupiec, J. A. (1998). High-spectral resolution data for monitoringScots pine (Pinus sylvestris L.) regeneration. International Journal of Remote Sensing,19(13), 2601–2608.

Sims, D. A., & Gamon, J. A. (2002). Relationships between leaf pigment content andspectral reflectance across a wide range of species, leaf structures and developmentalstages. Remote Sensing of Environment, 81, 337–354.

Spanos, I. A., Daskalakou, E. N., & Thanos, C. A. (2000). Postfire, natural regeneration ofPinus brutia forests in Thasos island, Greece. Acta Ocologica, 21(1), 13–20.

Thanos, C. A., & Marcou, S. (1991). Post-fire regeneration in Pinus brutia forest ecosystemsof Samos Island (Greece): 6 years after. Acta Oecologica, 12(5), 633–642.

Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoringvegetation. Remote Sensing of Environment, 8, 127–150.

Turner, M. G., Romme,W. H., Gardener, R. H., & Hargrove,W.W. (1997). Effects of fire sizeand pattern on early succession in Yellowstone National Park. Ecological Monographs,67(4), 411–443.

Turner, M., Romme, W. H., & Tinker, D. B. (2003). Surprises and lessons from the 1988Yellowstone fires. Frontiers in Ecology and the Environment, 1(7), 351–358.

Turner, M. G., Tinker, D. B., Romme, W. H., Kashian, D. M., & Litton, C. M. (2004). Landscapepatterns of sapling density, leaf area, and aboveground net primary production in post-fire lodgepole pine forests, Yellowstone National Park (USA). Ecosystems, 7, 751–775.

Vandvik, V., Heegaard, E., Maren, I. E., & Aarrestad, P. A. (2005). Managing heterogeneity:The importance of grazing and environmental variation on post-fire succession inheathlands. Journal of Applied Ecology, 42(1), 139–149.

Vogelmann, J. E., Rock, B. N., & Moss, D. M. (1993). Red-edge spectral measurementsfrom sugar maple leaves. International Journal of Remote Sensing, 14, 1563–1575.

Wu, W., & Rao, S. S. (2006). Fuzzy analysis of geometric tolerances using intervalmethods. Proceedings of the Institution of Mechanical Engineers, Part C: Journal ofMechanical Engineering Science, 220(4), 489–497.