Mapping fuels in Yosemite National Park

Embed Size (px)

Text of Mapping fuels in Yosemite National Park


    Mapping fuels in Yosemite National ParkSeth H. Peterson, Janet Franklin, Dar A. Roberts, and Jan W. van Wagtendonk

    Abstract: Decades of fire suppression have led to unnaturally large accumulations of fuel in some forest communities in thewestern United States, including those found in lower and midelevation forests in Yosemite National Park in California. Weemployed the Random Forests decision tree algorithm to predict fuel models as well as 1-h live and 1-, 10-, and 100-h dead fuelloads using a suite of climatic, topographic, remotely sensed, and burn history predictor variables. Climate variables andelevation consistently were most useful for predicting all types of fuels, but remotely sensed variables increased the kappaaccuracymetric by 5%12% age points in each case, demonstrating the utility of using disparate data sources in a topographicallydiverse region dominated by closed-canopy vegetation. Fire history information (time-since-fire) generally only increased kappaby 1% age point, and only for the largest fuel classes. The Random Forests models were applied to the spatial predictor layers toproduce maps of fuel models and fuel loads, and these showed that fuel loads are highest in the low-elevation forests that havebeen most affected by fire suppression impacting the natural fire regime.

    Rsum : La suppression du feu pendant plusieurs dcennies a entran d'importantes accumulations de combustibles danscertaines communauts forestires de l'Ouest des tats-Unis, incluant celles qu'on retrouve dans les forts situes a basse etmoyenne altitude dans le parc national de Yosemite en Californie. Nous avons utilis l'algorithme arborescent de dcision selonla mthode des forts alatoires pour laborer des modles de prdiction des combustibles aussi bien que des charges decombustibles vivants de 1 h etmorts de 1, 10 et 100 h a l'aide d'une suite de variables indpendantes climatiques, topographiques,obtenues par tldtection et portant sur l'historique des feux. Les variables climatiques et l'altitude taient invariablement lesplus utiles pour prdire tous les types de combustibles mais les variables obtenues par tldtection augmentaient la mtriquede prcision kappa de 5 a 12 points de pourcentage dans chaque cas, dmontrant l'utilit d'utiliser des sources disparates dedonnes dans une rgion o la topographie varie et qui est domine par un couvert forestier ferm. L'information concernantl'historique des feux (l'intervalle entre les feux) augmentait gnralement kappa de seulement 1 point de pourcentage et celaseulement pour les classes de combustibles les plus importantes. Les modles obtenus par la mthode des forts alatoires ontt appliqus aux couches de prdiction spatiale pour produire des cartes des modles de combustibles et des charges decombustibles qui montraient que les charges de combustibles sont les plus importantes dans les forts a basse altitude danslesquelles la suppression du feu a eu le plus d'effet en modifiant le rgime naturel des feux. [Traduit par la Rdaction]

    IntroductionFire is an integral part of ecosystems in the western United

    States. Decades of fire suppression have led to unnaturally largeaccumulations of fuel in some forest communities, includingthose found in lower and midelevation forests in Yosemite Na-tional Park (YNP) (Skinner and Chang 1996). This increasedamount of available fuel has likely contributed to a marked in-crease in burned area in the United States: 28 million ha burnedbetween 2000 and 2009 compared to 13 million ha burned in eachof the previous three decades ( Westerling et al. (2006) demonstrated that awarmer, drier climate cycle is also a likely factor in recent in-creases in area burned, especially for higher elevation forests. Inan effort to return fire to the YNP ecosystem, the park has per-formed prescribed fires since 1970 and allowed wildland fires toburn under prescribed conditions since 1972 (van Wagtendonkand Root 2003). Fuels have been sampled throughout the park,but spatial maps of fuels would aid in prioritizing areas in need offuel management activities, such as prescribed burning.Fuel models describe the amount and condition of the surface

    fuels (e.g., leaf and needle litter, fallen branch wood, and smalllive trees and shrubs) throughwhichmost fires burn (Albini 1976).They are used, in concert with information on weather and topog-raphy, to predict the growth of prescribed or wildland fires by fire

    spread models such as FARSITE (Finney 2004) or HFire (Petersonet al. 2009), allowing for fire risk assessment of active, prescribed,and modeled fires. Remote sensing provides the opportunity toefficiently develop maps of fuel models for large areas; however,surface fuels are not directly visible to remote sensing systems formost fuel models due to overstory vegetation (Keane et al. 2001).Hence, ancillary data describing site potential for growing vege-tation (i.e., surface fuels) can be incorporated to map fuels.A majority of the fuel studies using remote sensing have involved

    predicting and mapping fuel models via direct or indirect mapping(Keane et al. 2001). In indirect mapping, vegetation mapping is per-formed first, and then each vegetation class is assigned to a fuelmodel using a look-up table (e.g., Keane et al. 2000). However, avegetation class may be composed of multiple types of fuel depend-ing on vegetation condition and density and other factors, such asstand history, so this approach is not commonly used (Keane et al.2001). In direct mapping, the classification algorithm directly pre-dicts fuel models for each pixel (e.g., Riao et al. 2002).

    Accuracies (percent correct classification) when predicting fuelmodels using remotely sensed data have ranged from 50% to 85%,with kappa coefficients of agreement (Congalton 1991) rangingfrom 0.03 to 0.54 for studies in coniferous ecosystems (e.g., Keaneet al. 2000, 2002; van Wagtendonk and Root 2003; Rollins et al.2004; Falkowski et al. 2005) and from 0.54 to 0.79 for studies in

    Received 11 May 2012. Accepted 12 November 2012.

    S.H. Peterson and D.A. Roberts. Department of Geography, University of California at Santa Barbara, CA 93106, USA.J. Franklin. School of Geographical Sciences & Urban Planning, Arizona State University, Tempe, AZ 85287, USA.J.W. van Wagtendonk. US Geological Survey, Western Ecological Research Center, Yosemite Field Station, El Portal, CA 95318-0700, USA.

    Corresponding author: Seth H. Peterson (e-mail:


    Can. J. For. Res. 43: 717 (2013) Published at on 13 November 2012.


    . J. F

    or. R

    es. D



    d fr

    om w






    m b

    y U




    Y O

    F T



    IA o

    n 11



    r pe


    al u

    se o


    mailto: seth@geog.ucsb.edu

  • shrublands and woodlands (e.g., Riao et al. 2002; Poulos et al.2007; Poulos 2009) (Table 1). Kappa is an accuracy measure, rang-ing from0 to 1, that accounts for pixels that are classified correctlysimply by chance, making it more robust than percent correctlyclassified. The higher accuracies for the shrubland/woodlandstudiesmay have been achieved because the vegetation/fuel is notbeing obscured by tree canopies for these ecosystems. To ourknowledge, Rollins et al. (2004) provided the only previous predic-tions of surface fuel loads in forested ecosystems; surface fuel loadwas measured as a continuous variable but discretized into threeordinal categories (low, medium, and high) for model trainingand validation. Their accuracy was 52%, with a kappa of 0.20.We used the Random Forests (RF) algorithm (Breiman 2001)

    implemented in R (R Development Core Team 2010) for our fuelclassification analysis. A majority of fuel classification studies(e.g., Keane et al. 2002; Rollins et al. 2004; Falkowski et al. 2005;Poulos et al. 2007; Poulos 2009) have used classification trees (CT)(Breiman et al. 1984). RF have not yet been used in the fuels liter-ature, although they are in common use in other disciplines (e.g.,Nicodemus et al. 2010; Bi and Chung 2011). CT recursively dividethe data into more homogeneous groups of the dependent vari-able through binary splits of the explanatory variable that bestreduces deviance at each particular node. A weakness in this ap-proach is that each division only optimizes the classification ofthe two groups generated by the split it does not necessarilyoptimize overall classification accuracy. The series of splits makesup a single tree. The RF algorithm generates an ensemble of treeswhose predictions are averaged to determine the value for eachdata record, resulting inmore robust predictions. A large numberof different trees are generated by (1) evaluating a random subset ofthe explanatory variables to make any given split (the number ofpredictor variables evaluated at each split is equal to the squareroot of the number of predictor variables used in the model) and(2) using bootstrapped subsets of data to generate the trees andthe remaining data to evaluate them. The first step means thattrees are less likely to suffer from an optimal initial split thatmight have a detrimental effect on accuracy further down the treeand the second that accuracy reported by RF is based on indepen-dent subsamples, a further improvement over CT. Variable impor-tance is determined by permuting the explanatory variables andmeasuring the resulting effect on classification accuracy (Breiman2001). RF is a nonparametric model so the accuracy of the trees isnot affected bymulticollinearity in the predictor variables (Bi andChung 2011), which is important for this research, as certain var