Boun a Venture 2012

emer

ao Ca

imatgh-rein th

empirical-statistical models, each developed from basal temperature of snow measurements in winter and ground-

INT

EmphighmanHoelzle, 2001; Lugon and Delaloye, 2001; Gardaz, 1997;

on-forbles a

High-resolution maps of permafrost distribution at regional

PERMAFROST AND PERIGLACIAL PROCESSESPermafrost and Periglac. Process., 23: 5268 (2012)Published online in Wiley Online Library(wileyonlinelibrary.com) DOI: 10.1002/ppp.1733Hoelzle et al., 1999; Imhof et al., 2000; Hoelzle, 1992; King,1992; Dobinski, 1998), Scandinavia (e.g. Isaksen et al., 2002;Jeckel, 1988; degard et al., 1996), Japan (e.g. Ishikawaand Hirakawa, 2000) and North America (Lewkowiczand Ednie, 2004; Bonnaventure and Lewkowicz, 2008,2010; Lewkowicz and Bonnaventure, 2008). These modelshave the advantage of requiring limited eld data for theircreation, in contrast to process-oriented or numerical models(Riseborough et al., 2008). They can also be used to generate

scales are needed in order to establish and compare current per-mafrost conditions during climatic warming. Sub-arctic regionslocated in the discontinuous permafrost zone, such as withintheYukon and northern British Columbia, could be strongly af-fected by future climatic change (AICA, 2004; IPCC, 2007).Higher air temperatures or greater winter precipitationmay leadto the degradation of discontinuous permafrost which is rela-tively warm (close to 0C) and potentially thin (Romanovskyet al., 2010; Smith et al., 2010; Lewkowicz et al., 2011). Thismapdon

* Cophy,K7L

Copygenerate a map of permafrost probability over an area of almost 500 000 km2 between 59N and 65N. The result isbroadly similar to previous permafrost maps with an average permafrost probability of 58 per cent for the region as awhole. There are notable differences in detail, however, because the main predictive variable used in the local modelsis equivalent elevation, which incorporates the effects of gentle or inverted surface lapse rates in the forest zone. Mostof the region shows permafrost distribution patterns that are non-linear, resembling those from continental areas suchas Mongolia. Only the southwestern area shows a similar mountain permafrost distribution to that in the European Alpswith a well-dened lower limit and a linear increase in probability with elevation. The results of the modelling can bepresented on paper using traditional classications into permafrost zones but given the level of detail, they will be moreuseful as an interactive online map. Copyright 2012 John Wiley & Sons, Ltd.

KEY WORDS: mountain permafrost modelling; BTS; equivalent elevation; surface lapse rate; Yukon; British Columbia

RODUCTION

irical-statistical modelling of permafrost distribution atspatial resolution has been used in mountain areas iny locations, including central Europe (e.g. Gruber and

a single mountain range. This is because the variable relatiships seen in the empirical data tend to be valid onlyspecic geographic areas. Understanding how these variarelationships behave over a larger geographic region ilogical next step (Lewkowicz and Bonnaventure, 2008).truthing of frozen-ground presence in summer. The models were blended using a distance-decay power approach toA Permafrost Probability Model for thColumbia, Canada

Philip P. Bonnaventure,1,2* Antoni G. Lewkowicz,1 Marian Kre

1 Department of Geography, University of Ottawa, Ottawa, Canad2 Department of Geography, Queens University, Kingston, Ontari

ABSTRACT

Permafrost maps are needed for infrastructure planning, clbut often lack sufcient detail for these purposes. The hisouthern Yukon and northern British Columbia presenteds at an unprecedentedly ne resolution, but this has beene almost exclusively for relatively restricted areas within

rrespondence to: Philip P. Bonnaventure, Department of Geogra-Queens University, 68 University Avenue, Kingston, Ontario3N6, Canada. E-mail: [email protected]

right 2012 John Wiley & Sons, Ltd.Southern Yukon and Northern British

1 and Michael C. Sawada1

nada

ic change adaptation strategies and northern developmentsolution (30 x 30m grid cells) probability model for theis paper (regional model) is a combination of seven localthawing could have potentially problematic consequences withrespect to infrastructure, geohazards and northern develop-ment, as well as potential greenhouse gas release (Brouchkovand Fukuda, 2002; Kneisel et al., 2007).Empirical-statistical permafrost probability models have

been generated for parts of the Yukon and northernBritish Columbia using the basal temperature of snow

Received 17 January 2011Revised 3 January 2012

Accepted 2 February 2012

(BTS) method and extensive ground-truthing of permafrostpresence or absence for several areas over a range ofclimatic zones (Lewkowicz and Ednie, 2004; Lewkowiczand Bonnaventure, 2008; Bonnaventure and Lewkowicz,submitted). The objective of this paper is to use theselocalised results to assemble a model of permafrostprobability at high resolution for the entire southern halfof the Yukon and northern-most British Columbia, an areaof almost 500 000 km2.

STUDY REGION

The study region covers 59N to 65N from 141W to as fareast as 1235W (Figure 1). This region is considered to bepart of the Cordilleran orogen geological grouping, compris-ing large mountain belts of deformed and metamorphosedsedimentary and volcanic rocks, mainly of the Phanerozoicand Proterozoic ages (Department of Energy, Mines andResources, Canada, 1974; Wahl et al., 1987; Eyles and Miall,2007). During the Wisconsinan glacial maximum, the south-ern and central portions of the study area were covered bythe thick ice masses of the Cordilleran ice sheet whereasthe northwestern portion around the Dawson area wasunglaciated (Duk-Rodkin, 1996).Elevations range from 250m asl in the Yukon River

1995). The entire region is in the Boreal Cordillera eco-zone, which is characterised by mountain ranges withnumerous lofty peaks and extensive plateaus with long,cold winters and short, warm summers, varying with el-evation and mountainside orientation (Natural ResourcesCanada, 2010).Climatic gradients across the region are relatively gen-

tle except in the southwest where there is a strong pre-cipitation gradient from 1415mm at Pleasant Camp to270mm at Whitehorse over a distance of 150 km. Conti-nentality increases northward and eastward. Annual pre-cipitation (outside St Elias) ranges between about 300and 400mm with the highest amounts in the Liard Basinclimatic region (Wahl et al., 1987). The entire study re-gion experiences winter-time inversions in surface lapserates (SLRs) through the forested zone (Lewkowicz andBonnaventure, 2011). The results of inverted SLRs onan annual basis are increased permafrost probabilities invalley bottoms in areas of high continentality while inmore maritime environments SLRs are gentle but normal,so that permafrost is less common at low elevations. Thechanges in SLR at treeline have been observed throughoutthe study region over multiple years but the reasons for thesechanges remain unclear (Lewkowicz and Bonnaventure,2011).Vegetation in the region comprises boreal forest with

to perD =Dnt Ca

A Regional Permafrost Probability Model 53Figure 1 Map of the study region showing modelling locations in relation1987). JC= Johnsons Crossing; SDH=Sa Dena Hes; F = Faro; K =Keno;Whitehorse, YT is located approximately 20 km northeast of WC and Pleasavalley to greater than 5000m asl in the St Elias Moun-tains. The region encompasses all terrestrial permafrostzones, from isolated patches in the southwest to continu-ous in the most northerly areas (Heginbottom et al.,colour online at wileyonlinel

Copyright 2012 John Wiley & Sons, Ltd.coniferous trees and some boreal broadleaf trees in lowlandareas, while sub-alpine forest, shrubs, alpine tundra, barrenpatches and exposed rock occur progressively at higherelevations (Wahl et al., 1987; Kremer et al., 2011). The

mafrost zones (Heginbottom et al. 1995) and climatic regions (Wahl et al.,awson; RR=Ruby Range; HS=Haines Summit; WC=Wolf Creek. Note:mp, BC is located approximately km south of HS. This gure is available in

ibrary.com/journal/ppp

Permafrost and Periglac. Process., 23: 5268 (2012)

54 P. P. Bonnaventure et al.northern-most portion of the study area is very close to anecosystem boundary where vegetation begins to transitionto arctic tundra with alpine sedges, grasses and shrubsdominating (Wahl et al., 1987).

Figure 2 (A) Fourth-order polynomial trend surface representing treeline; (B) thmap displaying the per cent of each individual model contributing to the regional mmodelling process outside of the intensive study areas. The intensive study are

truncated at 800m and 1600m in (A); and surface lapse rate tren

Copyright 2012 John Wiley & Sons, Ltd.Climatic regions in relation to the individual study areas areshown in Figure 1 and described more fully in previouspublications (Lewkowicz and Ednie, 2004; Lewkowicz andBonnaventure, 2008;Bonnaventure andLewkowicz, submitted).

ird-order polynomial trend surface representing surface lapse rate; and (C)odel at eight locations selected to demonstrate the operation of the blendedas are shown in black outline around each area. Treeline trend surface isd surface is truncated at +1C/km and 6.5C/km in (B).


A Regional Permafrost Probability Model 55METHODS

Data Sources for the Regional Model

This section explains the sources of the information required aswell as the steps followed during data processing. The nalassembly of the regional model involved map algebra usingArcGIS with multiple steps completed in the Raster Calculator.

Field Data Collection.Individual local empirical-statistical permafrost probability

models were generated for seven eld areas ranging in sizefrom 200 to 1400 km2: Wolf Creek, Ruby Range, HainesSummit, Johnsons Crossing, Faro, Keno and Dawson(Figures 1 and 2). Local models were based on a combinationof BTS measurements in winter and ground-truthing insummer to verify directly the presence or absence of frozenground. Logistic regression was then used to relate predictedBTS temperatures to the ground-truthing observations.Full descriptions of the methods have been outlined in previ-ous publications (Lewkowicz and Ednie, 2004; Bonnaventureand Lewkowicz, 2008, submitted).

Digital Elevation Model.A digital elevation model (DEM) was needed to allow

predictions to be made for the entire region. The Yukonportion of the DEM was provided by the Yukon GeologicalSurvey (Geomatics Yukon, 2006) with a resolution of 30 x30m. Sections at the same resolution were added forparts of British Columbia to the south and the NorthwestTerritories to the east (both provided by Geobase, 2010)and Alaska to the west (provided by the United StatesGeological Survey, 2010). These outer DEM sections wereincluded to extend the modelled area into northern BritishColumbia and to eliminate errors associated with edgeeffects during analysis. They were incorporated into theYukon DEM using the Mosaic tool in ArcGIS.

Solar Radiation Modelling.The solar modelling processes were essentially the same as

those described in Bonnaventure and Lewkowicz (2008,2010). However, the creation of a potential incoming solarradiation (PISR) model for the entire study region that tookinto account latitude changes represented a considerablechallenge. As a result of software restrictions in the ArcGISArea Solar Radiation tool, relatively small sections of thePISR model had to be run individually. The total area wasthus divided into 34 sections of the DEM that were 0.5 in lat-itude, and varied in longitude from about 0.5 to about 2.These sections were overlapped by 0.125 in both latitudeand longitude to eliminate any errors associated with edgeeffects relating to topography and shading.The climatic inputs needed to run the Area Solar Radiation

tool were derived from: (1) air and ground temperatures;(2) light-intensity logger data from our logger network(Lewkowicz and Bonnaventure, 2011); and (3) cloud cover

data from nearby Environment Canada stations. The cloud

Copyright 2012 John Wiley & Sons, Ltd.cover percentages during the snow-free period were classi-ed daily as 0, 50 or 100 per cent. The percentage wasassigned by examining the offset between air and groundtemperatures (Bonnaventure and Lewkowicz, 2008). Whenground temperatures rise more rapidly than air tempera-tures during the day this reects clear conditions, whereasthe opposite reects overcast conditions. These interpreta-tions were compared to the light-intensity logger data aswell as the hourly cloud cover percentages from the nearestmeteorological station. The amount of cloud cover (diffu-sivity) for each of the 34 DEM sections was set to one ofthree values based on the results of the analyses (65, 70 or75%).The snow-free period, which is needed to decide on

the duration of the PISR calculations, was established byexamining logger data from all of the study areas to deter-mine typical dates when snow covered the surface (autumn)and when ground temperatures rose above 0C (spring). Thesnow-free period varies within the region and from year toyear, and is partially controlled by elevation and local sitecharacteristics. For the purpose of this model, however, itwas set from 15 May to 30 September. The highest sitesmay have a snow-free season that is slightly shorter thanthat used (i.e. PISR values may be over-predicted), but thishas a limited impact on the nal outputs of the permafrostmodel as these areas already exhibit high probabilities.Once all 34 PISR models had been created, they were

assembled into one grid using the Mosaic tool in ArcGIS.Modelled PISR is in MJ/m2 with the highest values beingpredicted for south-facing, high-elevation slopes, and thelowest values for steep, low-elevation, north-facing slopes.

Equivalent Elevation.The relationship between permafrost distribution and

elevation is complex in the region and valley bottoms arealso at relatively high elevations (typically 300700m asl)where permafrost may be present. The concept of equivalentelevation (Lewkowicz and Bonnaventure, 2011) was devel-oped to deal with these complexities. True elevations in aDEM are adjusted to reect mean annual air temperatures(MAATs) based on local SLRs. The numerical elevationsof grid cells below treeline are changed to take into accountweakened or reversed SLRs in the forest compared to thestrong normal negative SLRs above treeline. Thus grid cellsthat are well below treeline may be increased signicantlyin elevation, areas close to treeline are changed very littleand areas above treeline remain unchanged. Equivalentelevation is evaluated numerically from (Lewkowicz andBonnaventure, 2011):

Z x Zt Zt Zx L1L2

(1)

where Zx is equivalent elevation (m asl), Zt is the elevation oftreeline (m asl), Zx is the actual grid cell elevation (m asl), L1is the measured or predicted SLR below treeline (C/km) and

L2 is the SLR above treeline (assumed to be 6.5 C/km).


Figure 3 Regional model permafrost probability map with study area boundaries and locations of Environment Canada stations.

56 P. P. Bonnaventure et al.

Copyright 2012 John Wiley & Sons, Ltd. Permafrost and Periglac. Process., 23: 5268 (2012)

The derivation of equivalent elevation for the entire studyregion required trend surfaces to represent the treeline andSLR for each grid cell (Lewkowicz and Bonnaventure,2011). The polynomial order of the trend surfaces wasselected based on goodness-of-t relative to the source data,as well as visual inspection, especially near the edges of the

grid for each model. The prediction for the region as a wholefor each permafrost probability model was multiplied by itscorresponding per cent power. These seven grids were addedtogether to produce the nal values for the regional modelwhich was tted by a series of masks that enabled blendingoutside the specic study areas while maintaining 100 percent of the inuence of the specic model inside.

RESULTS

Regional Model Predictions

The regional model produced a probability of permafrostdistribution grid with an average value of 58 per cent forthe entire study region excluding currently glaciated terrain(Figure 3). This prediction is for near-surface permafrostwhich can be detected with the BTS and ground-truthingmethods that were used in the eld.

c portions of the region

Mid North Extreme southwest

Moderate High NoneModerate Low NoneHigh High High

A Regional Permafrost Probability Model 57study area. To predict treeline, a fourth-order polynomial trendsurface was created in ArcGIS using a series of 43 points wheretreeline elevation had been inspected (Figure 2A). This in-volved establishing the average value of treeline elevation us-ing National Topographic System maps for each of the maptiles in the study area. These values were checked by extensivevisual inspection of Google Earth images. A third-order poly-nomial was used for SLR (Figure 2B). SLRs in the Yukonare highly related to the annual range of monthly air temper-atures (i.e. July mean minus January mean). This relationshipcan be expressed in a linear regression model (r2 = 0.96) andused to predict the SLR for the entire region (Lewkowiczand Bonnaventure, 2011). Larger amplitudes produce SLRsthat are inverted or close to 0C/km, whereas smaller am-plitude sites have normal negative SLRs (Lewkowicz andBonnaventure, 2011). Using these correlations, measured SLRsat seven sites and calculated values for a further 10 locations,were used to create this trend surface.Equation (1) was used with the treeline and SLR trend

surfaces and the DEM to calculate equivalent elevation foreach grid cell in the entire study area (Lewkowicz andBonnaventure, 2011).

Regional Permafrost Probability Model

Each of the individual local models was used with theregional DEM (slope grid), the regional PISR grid and theregional equivalent elevation grid to make permafrost probabil-ity predictions for the entire area. The regional model wasassembled from these predictions using a distance-decay func-tion to weigh the contribution of each of the seven individualmodels to a given grid cell and then assigning a permafrostprobability value based on a sum of scores (Figure 2C).The decay grid for each model was calculated using the

Euclidian Distance tool in ArcGIS. The grid originated fromthe outline of each set of study area boundaries, which werecreated using a 10-km buffer around all of the eld samplepoints and dissolving the buffers together. Based on thespacing of the sample areas, it was decided that each modelwould have 100 per cent of its inuence within its ownstudy area and then decay at a rate dened by:

Y e0:01x (2)

Table 1 Permafrost probability gradient information for speci

Locations South

Valleys below treeline LowElevations close to treeline ModerateElevations above treeline HighCopyright 2012 John Wiley & Sons, Ltd.where x is measured in kilometres. This produced an indi-vidual decay grid in which a models inuence was halvedat a distance of approximately 100 km.The distance-decay grids for each area were summed to

obtain a sum of powers grid. This was used as a denominatorfor each individual grid in order to obtain a per cent power

Figure 4 Observed presence vs predicted probability of permafrost atground-truthing sites (n = 771). Note: Points for predicted probability areplotted at the centre of each 0.1 division of values. Thick line is 1:1 line.Labels represent the number of eld observations in each range. This gureis available in colour online at wileyonlinelibrary.com/journal/pppPermafrost and Periglac. Process., 23: 5268 (2012)

58 P. P. Bonnaventure et al.Figure 3 shows several geographic gradients in permafrostprobabilities, which are summarised in Table 1. Permafrostprobability follows these patterns largely because of differ-ences in SLRs, and this results in the interngering ofpermafrost at low and high elevations. Only in the extremesouthwest of the study region, where temperatures arerelatively high, the snowpack is several metres thick and thereis a well-dened lower limit of permafrost at elevations abovetreeline, does the permafrost distribution pattern correlatedirectly with elevation and therefore resembles that of theEuropean Alps and southern Norway.

Figure 5 Comparison of predicted permafrost probabilities for: (A) known roclassied rock glaciers (n = 225, active and inactive) (Page, 2009). Labels represen

colour online at wileyonlinel

Copyright 2012 John Wiley & Sons, Ltd.Regional Model Comparisons

Validation of the regional model would require an indepen-dent and spatially distributed set of data for permafrostpresence and absence. Such a data-set does not exist.However, it is possible to compare the regional modelpredictions with the more than 750 observations of perma-frost presence or absence made in the individual study areas.Fully independent comparisons can be also made with thedistribution of rock glaciers in the region, with results from anumber of ground temperature-monitoring boreholes, and with

ck glaciers sites (n = 1675) in the southern Yukon (Page, 2009); and (B)t percentage of features in each probability range. This gure is available inibrary.com/journal/ppp


Table 2 Regional model predictions for borehole sites ordered from north to south (Source: Global Terrestrial Network forPermafrost, 2011)

Borehole locationElevation(m asl)

Latitude/Longitude

Permafrostpresent

MAGT (C)/Deptha (m)

Regionalmodel

probability

Red Creek 780 65 9 N/138 22W Yes 2.1/7.0 0.72Eagle Alaska, USA 269 64 46N/141 10W Yes 2.4/20.0 0.79Dawson 320 64 04N/139 26W Yes 2.0/3.8 0.72Sixty Mile 1462 63 54N/140 48W Yes 3.6/23.0 0.97Mayo 504 63 36N/135 52W Yes 1.1/16.0 0.69Carmacks 874 62 19N/136 40W No +0.3/22.5 0.70Ross River 659 61 58N/132 27W Yes 0.1/9.0 0.63Alpine Burwash 1842 61 27N/139 24W Yes 2.0/30.0 0.98Tuchitua km 161 1220 61 19N/129 36W Yes 0.5/3.2 0.66Takhini Valley 1200 60 54N/135 5 W Yes 0.6/5.0 0.21Mount McIntyre 1570 60 37N/135 8 W Yes 1.6/20.0 0.80Mount Granger 2066 60 32N/135 15W Yes 5.4/0.5 0.99Wolf Creek palsa 25 1260 60 29N/135 13W Yes 0.3/5.0 0.33Wolf Creek Geological Survey of Canada (GSC) 1195 60 28N/135 11W Yes 0.5/15.3 0.27Alaska Highway 844 677 60 28N/133 29W Yes 0.1/1.0 0.02Alaska Highway 825 723 60 22N/133 6 W Yes 0.1/4.0 0.13Northern Dancer 1623 60 0 N/13136W No +0.4/20.0 0.95aMean annual ground temperature (MAGT) is recorded at the depth indicated.

Figure 6 Regional permafrost model predictions for borehole sites in relation to measured ground temperatures. AH=Alaska Highway; WC=Wolf Creek.This gure is available in colour online at wileyonlinelibrary.com/journal/ppp

A Regional Permafrost Probability Model 59


the predictions from the Sa Dena Hes (SDH) area, an areamodelled within the study region but not used in thecreation of the regional model (Bonnaventure and Lewkowicz,submitted).

Ground-truthing Points.The comparison of the presence or absence of permafrost

at individual sites grouped by predicted probability from theregional model indicates a good t (Figure 4). However, all

the re

60 P. P. Bonnaventure et al.Figure 7 Comparison of (A) the Sa Dena Hes local model (LMSDH) to (B)

This gure is available in colour online at

Copyright 2012 John Wiley & Sons, Ltd.gional model for the SDH area and (C) the difference between (A) and (B).

wileyonlinelibrary.com/journal/ppp


data except the SDH results (approximately 90 points) wereused to generate the model so it is not surprising that thebroad patterns are represented. The comparison suggeststhat the regional model may slightly over-predict permafrostprobabilities in areas that are classied as extensive orcontinuous permafrost, while slightly under-predictingprobabilities in areas of sporadic discontinuous permafrost.Overall, the predicted trends in probability reect theobserved patterns within 10 per cent.

Rock Glaciers.A database of rock glaciers created for the Yukon (Page,

2009) contains the locations of all known features within thesouthern Yukon as well as a selection that were classiedbased on activity. Active and inactive rock glaciers are pointindicators of permafrost, while relic rock glaciers should beindicative of permafrost absence (Haeberli et al., 2006;Janke, 2004, 2005). Almost 90 per cent of the rock glacierlocations in the complete database (n= 1675) occur at siteswhere the predicted probability of permafrost is 0.8 or greater(Figure 5A). Only 1 per cent of the rock glaciers are at siteswith permafrost probabilities less than 0.5, while 28 per centof the regions Yukon area falls into this category. In thesub-sample of classied rock glaciers (n = 225: active

n = 191, inactive n = 34), more than 85 per cent of the activefeatures are in locations with probabilities greater than 0.8,while ~ 70 per cent of the inactive features are in this category(Figure 5B). Convective heat exchange in rock glaciers canhelp create colder ground conditions than on adjacent slopes(Haeberli et al., 2006; Guodong et al., 2007) and those havenot been taken into account in the regional model. Therefore,the overall pattern in which rock glaciers are mainly present inareas with high permafrost probability supports the validity ofthe modelling process.

Borehole Temperatures.Ground temperatures have been recorded at almost 20

borehole or other monitoring sites in the region (Table 2). Be-cause most of these sites were instrumented with the goal ofstudying permafrost, the great majority exhibit perenniallyfrozen ground. There is a relationship between predicted prob-ability and most measured ground temperatures (Figure 6),with the latter being generally lower at high probabilities.However, permafrost can be present at sites which have verylow predicted probabilities, and conversely, may be absent atsites with high predicted probabilities.The two non-permafrost sites are Northern Dancer (pre-

dicted probability 0.95) and Carmacks (predicted probability

Hescal m

A Regional Permafrost Probability Model 61Figure 8 (A) Predicted probabilities at ground-truthing sites in the Sa Dena(y-axis) of permafrost probabilities (x-axis) from the regional and SDH lojournal/p

Copyright 2012 John Wiley & Sons, Ltd.(SDH) area from the regional and local models. (B) Frequency distributionodels. This gure is available in colour online at wileyonlinelibrary.com/

pp


0.7). At the Northern Dancer site, a possible explanation ofpermafrost absence is that the model is not weighting thePISR variable as heavily as it should. Northern Dancer islocated at a high elevation, highly exposed south-facing slope,which receives abundant insolation. A second possibility isthat site disturbance has caused thaw since the borehole wasdrilled in 1978. The organic mat had been completely strippeddown to the bedrock prior to the borehole being drilled, andthe site was used as a mining road. The Carmacks boreholeis also in a highly disturbed area that is used as a mining road.Thus it is possible that anthropogenic factors could have led tothe loss of permafrost at these locations, which is somethingthe model cannot accommodate. However, given the natureof probability modelling it is also true that high probabilitiesdo not equate denitively to permafrost presence at any givensite, but indicate only that it is likely to be present.

SDH Model.The regional model was created without incorporating the

results from the SDH area (Figure 1), because modelling forthis area employed the normalised difference vegetation in-dex (NDVI) as one of the inputs. This variable could not begenerated for the entire study region at the scale needed.Consequently, results from the local SDH model can beindependently compared with those of the regional model(Figure 7). The overall percentages of areas predicted to

be underlain by permafrost are similar, with 30 per centfor the local model and 39 per cent for the regional one.In detail, the regional model indicates a greater inuenceof elevation whereas the SDH local model is based onlyon slope and NDVI. For most of the lowland areas,however, the two models predict comparably low to interme-diate probabilities (0.20.3). A comparison of the ground-truthing results for the area with the modelled permafrostprobabilities from both the SDH and the regional modelshows that the latter is better at discriminating betweenpermafrost and non-permafrost sites (Figure 8A). The greatmajority of sites with permafrost are in locations where theprobability is greater than 0.5, and most of the non-permafrostsites have probabilities less than 0.5. The local SDHmodel, on the other hand, produced predictions in a rangeof probabilities from 0.2 to 0.5. The predicted values arehigher for the regional model than the local one for thearea as a whole (Figure 8B) and we believe these to bemore realistic.

DISCUSSION

Comparison with Previous Maps

The regional model is the rst attempt internationally tomodel permafrost using empirical-statistical methods for

ated p

62 P. P. Bonnaventure et al.Figure 9 Regional model classied into traditional permafrost classes of isol

(5090%) and continuous (> 90%). Permafrost map of C

Copyright 2012 John Wiley & Sons, Ltd.atches (< 10%), sporadic discontinuous (1050%), extensive discontinuous

anada boundaries after Heginbottom et al. (1995).


Another potential source of error in the regional model isthe trend surfaces which were needed to predict equivalentelevation values for each grid cell. Figure 10 shows thecomparison between observed values and those predictedfrom the trend surface. For the treeline surface, more thanhalf of the 43 points were within 20m of the measuredvalue and the maximum deviation was approximately

Figure 10 Comparison of: (A) observed and predicted treeline fromfourth-order polynomial trend surface; and (B) observed or predicted sur-face lapse rates (SLR) (from annual amplitude of monthly air temperatures)with values predicted from third-order polynomial trend surface. This g-ure is available in colour online at wileyonlinelibrary.com/journal/ppp

A Regional Permafrost Probability Model 63an area this large and at such a ne resolution. Conse-quently, it is not possible to directly compare the resultswith previous maps produced with a coarser resolution.However, it is possible to classify the predictions into prob-ability classes that conform with previous maps andexamine these at a small scale (i.e. large area) (Figure 9).The regional model is similar to the permafrost map ofCanada (Heginbottom et al., 1995) but contains orders ofmagnitude more detail. The greatest differences are wherethe regional model classies areas of high elevation in thesouthern part of the region as continuous permafrost. Themap by Heginbottom et al. (1995) apparently did notdirectly take elevation sufciently into account, and focusedmore on geology and coarse latitudinal and climatic gra-dients. It is also evident that there are areas at the north-ern end of the study area that are classied as sporadicin the regional model but as extensive on the map byHeginbottom et al. (1995). These are the result of annualinversions in SLRs that are present in the highly conti-nental Dawson and Keno areas in the northern parts ofthe study region (Figure 1). Another difference is theidentication of an area of continuous permafrost aroundBeaver Creek. This location has a MAAT around5.5C which would place it close to the assumed tem-perature under which continuous permafrost can form(6C) (French, 2007). Although an area of continuouspermafrost is not shown at this scale, Heginbottom andRadburn (1992) produced a 1:1 000 000-scale map onwhich higher elevation terrain in this area is categorisedas continuous, while our predicted SLRs are near zero inthis area meaning that permafrost probability is actuallyindependent of elevation. Finally, the zone of extensivediscontinuous permafrost in the regional model is largerthan in the map by Heginbottom et al. (1995), especiallyin the area to the south of Faro. Since there was noground-truthing undertaken there, it is possible thatthe regional model is over-predicting for this area wheremultiple local models contribute (Figure 2).

Inaccuracies and Uncertainties in the Regional Model

Inaccuracies and uncertainties in the regional modelrelate to the derivation of the local models and to themethods used to extend the modelling across such alarge region.Sources of uncertainty exist in relation to the BTS model-

ling for the individual areas themselves. These mainly relateto snowpack development at individual sites, samplingconstrained by logistics and inter-annual variation in winterconditions, particularly air temperatures. They have beendiscussed extensively in the literature (Brenning et al.,2005; Gruber and Hoelzle, 2001) as well as in our previouspublications (Lewkowicz and Ednie, 2004; Bonnaventureand Lewkowicz, 2008, submitted). Our logistic regressionmodelling differs from that undertaken elsewhere, however,through the use of ground-truthing to constrain the pre-dictions from the BTS measurements and this provides

additional condence in the results (see Figure 4).

Copyright 2012 John Wiley & Sons, Ltd.120m asl (Figure 10A). For the SLR trend surface, more


rface

64 P. P. Bonnaventure et al.than half of the 17 points were within 0.2C/km of thetrend surface prediction and the maximum deviations were+1.2 and 1.4C/km (Figure 10B).The deviations of the trend surfaces from the measured

values affect the permafrost probability results in a numberof ways that depend on actual values and whether the trendsurface is over- or under-predicting (Table 3). Spatially, thegreatest potential sources of error are associated with the tree-

Table 3 Implications and results of inaccuracies in the trend su

Error in trend surface: meaning

Treeline elevation over-predicted: model has placed trees abovethe position of treeline

Treeline elevation under-predicted: model has placed tundraareas within a forested environment

SLR over-predicted: the SLR is too positive so that invertedSLRs are larger or normal SLRs are gentler

SLR under-predicted: the SLR is too negative so that invertedSLRs are too small or normal SLRs are enhanced

SLRs = Surface lapse rates.line trend surface in southwest areas around the ice elds(Figure 11A). The ice elds themselves were excluded fromanalysis because the interaction between glaciers andpermafrost was not considered in the model. In the SLRsurface, the largest errors are present around the PellyRanch area (SLR too negative) and the Faro area (SLRtoo positive) (Figure 11B). Around Faro, permafrostprobability is therefore over-predicted in valley bottomsand in the Pelly Ranch area, permafrost probability isunder-predicted in valley bottoms (Table 3). In the PellyRanch area, however, the relief is relatively subdued whichlimits the impacts.The distance-decay function used in the regional model is

an additional source of uncertainty (Equation (2)). Very littleliterature exists on the interchangeability of permafrostmodels among areas (Lewkowicz and Bonnaventure, 2008).Figure 1 shows that the study area covers six different climaticzones (Wahl et al., 1987), and this was considered whenchoosing the decay function, but it was not deemed the mostimportant factor. The inuence of each model had to beextended so that sites close to an individual study area werealmost wholly represented by that model. Tests with a lineardecay function, however, produced highly blended resultsacross the region and especially where study areas are closeto each other (Wolf Creek, Johnsons Crossing and Haines

Copyright 2012 John Wiley & Sons, Ltd.Summit). The exponential decay function (Equation (2))resolved this problem with an exponent chosen based on themacro-topography, the gentle climatic gradients (except inthe southwest) and our judgement of what appeared reason-able in terms of model weightings across the region (seeFigure 2).

Regional Model Use

modelling.

Impact to the regional model

Permafrost probability is under-predicted at higher elevationsespecially immediately above the true treeline. This is becausethe normal SLR of 6.5C/km is used above treeline. Areasthat have inverted SLRs are affected close to treeline andinvalley bottomsPermafrost probability is over-predicted in the uppermost partof the forest. Areas that have inverted SLRs are most affectedby this errorPermafrost probability is over-predicted in valley-bottomlocations but the impact is greater for inverted SLRs becausepermafrost probabilities are high in valley-bottom locations.These impacts are less for locations with normal SLRs wherevalley bottoms have low probabilitiesPermafrost probability is under-predicted in valley-bottomlocations but the impact is greater for inverted SLRs becausepermafrost probabilities are high in valley-bottom locations.These impacts are less for locations with normal SLRs wherevalley bottoms have low probabilitiesCurrent permafrost distribution maps t into one of twogroups which can be differentiated based on the size of areathey cover. Small-scale maps cover very large areas (e.g.Heginbottom et al., 1995, 1:7 500 000). These maps areuseful for examining trends in permafrost at the continentalor polar level but have limited practical use (e.g. engineeringprojects or geohazard planning) especially in discontinuouspermafrost zones. Large-scale maps cover small areas at aner resolution (e.g. Lewkowicz and Ednie, 2004, presentedat 1:200 000 but with 30 x 30m resolution), show detailedpermafrost distributions and account for topographic effects,but are very limited in areal extent.Given the areal coverage of the regional model, the most

signicant comparable permafrost maps include the perma-frost map of Canada (Heginbottom et al., 1995), the perma-frost characteristics of Alaska map (Jorgenson et al., 2008)and the International Permafrost Associations circum-arcticmap of permafrost and ground ice conditions (Brown et al.,1997). The attributes of these maps are given in Table 4.The regional model represents a signicant improvement

on existing permafrost maps of the area (Heginbottom et al.,1995), providing much more detailed spatial information.This feature makes the regional permafrost probabilitymap useful at multiple scales, and it should therefore be


Figure 11 Fitted trend surfaces showing deviations from input values for (A) treeline elevation and (B) surface lapse rate.

A Regional Permafrost Probability Model 65


ea to

Perat

Permdistthicfeatandtem

Permdistdepfeat

Permdistground ice permafrost maps

such as Heginbottomincludes a relicpermafrost class

ual a

66 P. P. Bonnaventure et al.made accessible at multiple scales. We therefore plan tomake it available as an online interactive high-resolutionmap in the near future.The creation of the regional model has led to a better

Table 4 Attributes of major permafrost maps comparable in ar

Map LocationScale/

resolution

Permafrostmap of Canada(Heginbottomet al., 1995)

All of Canada 1:7 500 000/unspecied

Permafrostcharacteristicsof Alaska map(Jorgenson et al.,2008)

Alaska, USA 1:7 200 000/2 x 2 km

IPAs circum-arcticmap of permafrostand ground iceconditions(Brown et al.,1997)

Northernhemisphere(25 90N,180W180E)

1:10 000 000/unspecied

IPA= International Permafrost Association; MAAT=mean annunderstanding of the controlling factors on the distributionof permafrost in the Yukon. Mountain permafrost distribu-tion in the Yukon differs from that of the European Alpswhere permafrost is highly correlated with elevation andPISR in a linear relationship and only above treeline. Inthe Yukon, permafrost is observed above and below tree-line, with the relationship between permafrost and elevationbeing non-linear. Similar relationships have been observedin Mongolia where vegetation cover and topographicwetness are the main factors governing the existence ofpermafrost (Etzelmller et al., 2006). As a result of thisdifference, permafrost in northwest North America, andlikely other similar areas, cannot be modelled using thesame empirical-statistical variables as the Alps (e.g. Gruberand Hoelzle, 2001). The development of the equivalentelevation variable has shown that SLRs can be reduced oreven inverted, which greatly impacts the distribution ofpermafrost. Thus the regional model takes into accountthe topography and individual climatic characteristics ofthe areas, representing a major advance over previousmaps.The regional model is expected to be of direct use as a

benchmark for future permafrost studies in the Yukon andfor applications in infrastructure planning and hazardsassessment. It can also be used to model the potential effectsof climate change on permafrost at high resolution byaltering input variables that control MAATs and SLR such

Copyright 2012 John Wiley & Sons, Ltd.as equivalent elevation (e.g. Bonnaventure and Lewkowicz,2010; Janke, 2005). The regional model could be developedto examine geohazard risk as the climate changes. As anexample, it would be possible to overlay the areas where

et al. (1995)for Canada

ir temperature.the regional model.

mafrosttributes Methodology Boundary classes

afrostribution,kness,uresgroundperatures

Rules-basedapproach fromsurcial geologyand ground icecontent samples

Five categories:isolated patches(< 10%), sporadicdiscontinuous(1050%), extensivediscontinuous(5090%), continuous(> 90%) and sub-seapermafrost

afrostribution,ths andures

Rules-based modelfrom MAAT inputcombined withsurcial geology

Same as Heginbottomet al. (1995) exceptno sub-sea class

afrostribution,

Combination of allnationally created

Same as Heginbottomet al. (1995) andpermafrost probability is most sensitive to a given tempera-ture increase, on slopes with gradients exceeding thresholdvalues. This type of information would be of direct use toplanners and policy-makers, serving as a tool for climatechange adaptation strategies throughout the region.

CONCLUSIONS

The following conclusions can be reached as a result of thisresearch:

1. The regional model represents the rst attempt to map per-mafrost distribution in the southern half of Yukon andnorthern British Columbia since Brown et al. (1997) andis several orders of magnitude more detailed than previouswork. It covers an area of almost 500 000 km2, which is farlarger than that represented in any other empirical-statisti-cal permafrost model to date. Overall, 58 per cent of theregion is predicted to be underlain by permafrost.

2. Trends in permafrost distribution in forested areas varywith the degree of continentality across the region.Maritime sites have normal (negative) SLRs below andabove treeline so that permafrost probability increaseswith elevation. In continental sites, the SLR can beinverted through the forest zone, leading to signicantamounts of permafrost being present in valley bottoms


inonithranthtiolama

thalthn.oued

asede.af

inor

E

upeanceheAfl SAe

e Hpregaonste

gemb Aost diria apn MoProc.554007.Fitzh, Onhe Pn. Jo

istrinessePerm101(19972-Xhttp

ay 2010].2006.ode

geom

Hoelzle M, Wegmann M, Krummenacher B.

A Regional Permafrost Probability Model 67spatial distribution of mountain permafrostat three sites in northwest Canada. ClimaticChange 105: 293312. DOI 10.1007/s10584-010-9818-5

Bonnaventure PP, Lewkowicz AG. Submitted.Permafrost probability modelling above andbelow treeline, Yukon, Canada. ColdRegions Science and Technology.

Brenning A, Gruber S, Hoelzle M. 2005. Sam-pling and statistical analyses of BTS measure-ments. Permafrost and Periglacial Processes16: 111. DOI: 10.1002/ppp.541

Brouchkov A, Fukuda M. 2002. Preliminarymeasurements on methane content inpermafrost, central Yakutia, and someexperimental data. Permafrost and PeriglacialProcesses 13: 187197.DOI: 10.1002/ppp.422

Brown J, Ferrians Jr OJ, Heginbottom JS,Melnikov ES. 1997. Revised February2001. Circum-Arctic Map of Permafrostand Ground-ice Conditions. National Snowand Ice Data Center/World Data Center forGlaciology: Boulder, CO. Digital media.

Department of Energy Mines and Resources,Canada. 1974. Physiographic regions of

250 000.Etzelmller B, Heg

Frauenfelder R, KMountain permafrusing a multi-critegl area, northerand PeriglacialDOI: 10.1002/ppp

Eyles N, Miall A. 2Geologic Journey.Limited: Markham

French HM. 2007. Tment, Third EditioChichester.

Gardaz JM. 1997. Dpermafrost, FontaAlps, Switzerland.cial Processes 8:(SICI)1099-1530PPP241>3.0.CO;

Geobase. 2010.[Accessed on 4 M

Geomatics Yukon.Digital Elevation Min 2007). ftp://ftp.Yukon and surrounding Canadian territory. 30m [Accessed on 15 Oc

Copyright 2012 John Wiley & Sons, Ltd.30 Meter Yukonl (data le compiledaticsyukon.ca/DEMs/

1999. Miniature temperature dataloggersfor mapping and monitoring of permafrostin high mountain areas: rst experiencesESF, Sharkhuu N,, Goulden C. 2006.stribution modellingproach in the Hvs-ngolia. Permafrostesses 17: 91104.

Canada Rocks, Theenry and Whitesidetario.eriglacial Environ-hn Wiley and Sons:

bution of mountains Basin, Valaisianafrost and Perigla-105. DOI: 10.1002/01)8:1

Periglacial Processes 10: 113124. DOI:10.1002/(SICI)1099-1530(199904/06)10:2 3.0.CO;2-A

Imhof M, Pierrehumert G, Haeberli W, KienholzH. 2000. Permafrost investigation in theSchilthorn Massif, Bernese Alps, Switzer-land. Permafrost and Periglacial Processes11: 189206. DOI: 10.1002/1099-1530(200007/09)11:33.0.CO;2-N

IPCC. 2007. http://www.ipcc.ch/ipccreports/assessments-reports.htm [Accessed on 20October 2010].

Isaksen K, Hauck C, Gudevang E, degrdRS, Sollid JL. 2002. Mountain permafrostdistribution in Dovrefjell and Jotunheimen,

Permafrost, 29 June3 July 2008, Fairbanks,Alaska, USA.

King L. 1992. Prospecting and mappingof mountain permafrost and associatedphenomenon. Permafrost and PeriglacialProcesses 3: 7381. DOI: 10.1002/ppp.3430030204

Kneisel C, Rothenbhler C, Keller F, HaeberliW. 2007: Hazard assessment of potentialperiglacial debris ows based in GIS-basedspatial modelling and geophysical eldsurveys: a case study in the Swiss Alps.Permafrost and Periglacial Processes 18:259268. DOI: 10.1002/ppp.593

Kremer M, Lewkowicz AG, Bonnaventure PP,Sawada M. 2011. Utility of classication

Lugon R, Delaloye R. 2001. Modeling alpinepermafrost distribution, Val de Rechy,Valais Alps (Switzerland). Norsk geograskTidsskrift 55: 224229.

Natural Resources Canada. 2010. Forest Eco-systems of Canada. http://ecosys.c.scf.rncan.gc.ca/classication/classif08-eng.asp[Accessed on 7 June 2010].

degard RS, Isaksen K, Mastervik M, BilldalL, Engler M, Sollid JL. 1996. Comparisonof BTS and Landsat TM data fromJotunheimen, southern Norway. NorskGeogrask Tidsskrift 53: 226233.

Page A. 2009. A topographic and photogram-metric study of rock glaciers in the southernYukon Territory. Unpublished MSc thesis,

Riseborough D, Shiklomanov N, Etzelmuller

68 P. P. Bonnaventure et al.resistivity tomography data. Norsk Geogra-sk Tidsskrift 56: 122136. DOI: 10.1080/00291950 2760056459

Ishikawa M, Hirakawa N. 2000. Mountainpermafrost distribution based on BTSmeasurements and DC resistivity sound-ings in the Daisetu Moutains, Hokkaido,Japan. Permafrost and Periglacial Pro-cesses 11: 109123. DOI: 10.1002/1099-1530(200004/06)11:23.0.CO;2-O

Janke JR. 2004. The occurrence of alpinepermafrost in the Front Range of Colorado.Geomorphology 67: 375389.

Janke JR. 2005. Modeling past and futurealpine permafrost distribution in the Color-ado Front Range. Earth Surface Processesand Landforms 30: 14951508.

Jeckel PP. 1988. Permafrost and its altitudinalzonation in N. Lapland. In Proceedings ofthe Fifth International Conference onPermafrost, Kane D, Hinkel K, Alton T,Pedersen F, Boatwright S (eds). Trondheim.Tapir: Trondheim; 1: 332337.

Jorgenson T, Yoshikawa K, Kanevskiy M,Shur Y. 2008. Permafrost characteristics ofAlaska. In Conference Proceedings fromthe Ninth International Conference onCopyright 2012 John Wiley & Sons, Ltd.tion in mountain permafrost models,Yukon, Canada. Permafrost and Perigla-cial Processes 22: 163178. DOI: 10.1002/ppp.719.

Lewkowicz AG, Bonnaventure PP. 2008.Interchangeability of mountain permafrostprobability models, northwest Canada.Permafrost and Periglacial Processes 19:4962. DOI: 10.1002/ppp.612

Lewkowicz AG, Bonnaventure PP. 2011.Equivalent elevation: a new method to in-corporate variable lapse rates into mountainpermafrost modelling. Permafrost and Peri-glacial Processes 22: 153162. DOI. 10.1002/ppp.720

Lewkowicz AG, Ednie M. 2004. Probabilitymapping of mountain permafrost usingthe BTS method, Wolf Creek, YukonTerritory, Canada. Permafrost and Perigla-cial Processes 15: 6780. DOI: 10.1002/ppp.480

Lewkowicz AG, Etzelmller B, Smith SL.2011. Characteristics of discontinuouspermafrost from ground temperature mea-surements and electrical resistivity tomogra-phy, southern Yukon, Canada. Permafrostand Periglacial Processes 22: 320342.DOI: 10.1002/ppp.703PermaB, Gruber S, Mrchenko S. 2008. Recent ad-vance in permafrost modelling. Permafrostand Periglacial Processes 19: 137156.DOI: 10.1002/ppp.615

Romanovsky VE, Smith SL, Christiansen HH.2010. Permafrost thermal state in the polarnorthern hemisphere during the Interna-tional Polar Year 20072009: a synthesis.Permafrost and Periglacial Processes 21:106116. DOI: 10.1002/ppp.689

Smith SL, Romanovsky VE, Lewkowicz AG,Burn CR, Allard M, Clow GD, YoshikawaK, Throop J. 2010. Thermal state of perma-frost in North America: a contribution to theInternational Polar Year. Permafrost andPeriglacial Processes 21: 117135. DOI:10.1002/ppp.690

United States Geological Survey. 2010. http://data.geocomm.com/dem/ [Accessed on 7May 2010].

Wahl HE, Fraser DB, Harvey RC, MaxwellJB. 1987. Climate of Yukon. CanadianGovernment Publishing Centre: Ottawa,Ontario, Canada.southern Norway, based on BTS and DC and regression tree analyses and vegeta- Department of Geography, University ofOttawa.frost and Periglac. Process., 23: 5268 (2012)

Date post:	17-Dec-2015
Category:	Documents
Upload:	t-viorela-rafaela
View:	8 times
Download:	0 times

Boun a Venture 2012

Documents