Content uploaded by David McCaffrey
Author content
All content in this area was uploaded by David McCaffrey on Nov 01, 2017
Content may be subject to copyright.
Content uploaded by David McCaffrey
Author content
All content in this area was uploaded by David McCaffrey on Nov 01, 2017
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=ujrs20
Download by: [David McCaffrey] Date: 01 November 2017, At: 09:22
Canadian Journal of Remote Sensing
Journal canadien de télédétection
ISSN: 0703-8992 (Print) 1712-7971 (Online) Journal homepage: http://www.tandfonline.com/loi/ujrs20
Assessing Fractional Cover in the Alpine Treeline
Ecotone Using the WSL Monoplotting Tool and
Airborne Lidar
David R. McCaffrey & Chris Hopkinson
To cite this article: David R. McCaffrey & Chris Hopkinson (2017) Assessing Fractional Cover
in the Alpine Treeline Ecotone Using the WSL Monoplotting Tool and Airborne Lidar, Canadian
Journal of Remote Sensing, 43:5, 504-512, DOI: 10.1080/07038992.2017.1384309
To link to this article: http://dx.doi.org/10.1080/07038992.2017.1384309
Accepted author version posted online: 09
Oct 2017.
Published online: 09 Oct 2017.
Submit your article to this journal
Article views: 11
View related articles
View Crossmark data
CANADIAN JOURNAL OF REMOTE SENSING
, VOL. , NO. , –
https://doi.org/./..
Assessing Fractional Cover in the Alpine Treeline Ecotone Using the WSL
Monoplotting Tool and Airborne Lidar
David R. McCarey and Chris Hopkinson
Department of Geography, University of Lethbridge, Alberta Water & Environmental Science Building, University Drive, Lethbridge, AB TK
M, Canada
ARTICLE HISTORY
Received January
Accepted September
ABSTRACT
As forest cover in mountain areas impacts headwater properties like habitat extent and downstream
water resources, it is important to assess and understand the changes that occur across forest
transition zones like the alpine treeline ecotone (ATE). Such changes occur slowly and manifest at
decadal to century time scales; however, the present benchmark spatial resolution for land cover
analysis from oblique repeat photographs, 100 m, is insucient to analyze anticipated changes in ATE
over a century-scale. In this research note, fractional cover classication of the ATE is achieved through
oblique photography analysis with the WSL Monoplotting Tool. Seven oblique photographs of the
West Castle Watershed (Alberta, Canada), collected by the Mountain Legacy Project, were gridded
to a 20-m resolution and assigned canopy cover classes by manual interpretation. Four canopy cover
classes (i.e., no cover, low vegetation, partial canopy, full canopy) were compared to lidar-derived
fractional cover. The extraction of canopy cover information from oblique photography at a resolution
of 20 m introduces the ability to assess and quantify changes in the ATE using century-scale oblique
photographic records.
RÉSUMÉ
La couverture forestière dans les régions montagneuses a un impact sur les propriétés des têtes
de bassin versant, telles que l’étendue des habitats et les ressources hydriques en aval. Il est donc
important d’évaluer et de comprendre les changements qui surviennent dans les zones de transition
forestières telles que l’écotone forestier alpin (EFA). Ces changements se produisent lentement et se
manifestent sur une échelle de temps décennale ou centennale. Cependant, le point de référence
actuelle en matière de résolution spatiale pour l’analyse de couverture terrestre avec photographies
obliques à répétition, 100 m, est insusant pour mesurer les changements anticipés dans les EFA à
l’échelle centennale. Cet article présente une classication de couverture terrestre fractionnaire pro-
duite par l’analyse de photographies obliques avec le WLS Monoplotting Tool. Sept photographies
obliques du bassin versant West Castle (Alberta, Canada), acquises par le Mountain Legacy Project,
ont été maillées à une résolution de 20 m et assignées à une classe de couverture forestière suivant
une interprétation manuelle. Quatre classes de couverture forestière (aucune couverture, végétation
basse, canopée partielle, pleine canopée) ont été comparées à une classication fractionnaire dérivée
du lidar. L’extraction d’informationsur la couverture forestière de photographies obliques à une résolu-
tion de 20 m rend l’évaluation et la quantication des changements dans les EFA à l’échelle centennale
possible.
Introduction
Alpine treeline ecotone (ATE), the transition zone
between alpine tundra and closed canopy, is character-
ized by distinct patterns in forest structure; decreases
in mature tree height, stand density, and fractional
cover correlate with increases in elevation (Körner 2012;
Tranquill i ni 1979). Environmental factors inuencing
these forest structure patterns are varied (Case and Dun-
can 2014; Holtmeier and Broll 2005; Weiss et al. 2015),
and understanding the degree to which multi-scale inter-
actions between anthropogenic, orographic, and climatic
CONTACT David R. McCaffrey david.mccaffrey@uleth.ca
factors aect shifts in ATE has relied on observation
methods, such as repeated plot measurements (Weiss
et al. 2015), dendrochronology (Liang et al. 2014), remote
sensing (Coops et al. 2013),andhistoricrecords(e.g.,
oblique photography; Hagedorn et al. 2014).
ATE observation methods are limited in the temporal
and spatial extents at which they can resolve change
(Danby 2011). For example, forest structure parameters
that characterize ATE, such as fractional cover, can be
observed at a high spatial resolution (∼1m)usingair-
borne lidar, but few extended lidar records presently exist
Copyright © CA SI
Downloaded by [David McCaffrey] at 09:22 01 November 2017
CANADIAN JOURNAL OF REMOTE SENSING 505
for periods longer than 20 years. Conversely, oblique
photograph records of ATE exist for periods exceeding
100 years (e.g., Trant et al. 2015), but quantitative analysis
of century-scale land cover change using oblique pho-
tographs has been restricted to low spatial resolutions
(∼100 m; Stockdale et al. 2015). In the Rocky Mountains,
where advance of ATE species, such as subalpine r (Abies
lasiocarpa), has been observed at 0.28 m yr−1–0.62 m
yr−1(Bekker 2005; Luckman and Kavanagh 2000), the
abilitytodetectshiftsinATEusingrepeatobliquepho-
tography will depend on increasing the spatial resolution
of analysis above 100 m; for example, if ATE has advanced
on the order of ∼50 m in a century, in order to detect
change using a century scale photographic record, the
spatial resolution of analysis must be higher than 100 m.
This research note introduces a methodology that
assesses fractional cover in oblique photography, which
increases both the temporal period and spatial scale
of ATE observation when applied to historic repeat
photographs. Using a technique pioneered by Stockdale
et al. (2015), land cover in an oblique photograph can be
rasterized, enabling direct comparison between repeat
photographs. The present research is, in part, a replica-
tion of Stockdale et al. (2015), but distinguishes itself by
adding analysis, which is critical for the application of
the method to observations of ATE: (i)wedemonstrate
that the rasterization method can be executed at a spatial
scaleof20m;(ii) we intentionally selected oblique pho-
tographs with overlapping view elds to test correspon-
dence between images; (iii) we improve upon a general
linear model of spatial error in a manner which allows
selection of the image with lowest error in cases of multi-
pleobservation;and(iv) we validate manual classication
of canopy fractional cover from oblique photographs
against fractional cover measured by airborne lidar.
Quantitative area analysis of oblique photographs
Prior limitations of quantitative oblique photogra-
phy analysis have been overcome with the advent of
software like the WSL Monoplotting Tool (WSL-MT;
note: WSL is the German acronym for “Eidgenössische
Forschungsanstalt für Wald, Schnee und Landschaft,”
the Swiss Federal Institute for Forest, Snow and Land-
scape Research), which produces georeferenced vector
data from oblique photographs using tie points between
oblique and aerial photographs, and high resolution
topographic data (Bozzini et al. 2012). Previous analyses
of oblique photographs to assess areas of change in the
ATE were mostly qualitative, relying on descriptions
of land cover change and estimations of spatial extent
(Butler and DeChano 2001;KlasnerandFagre2002;
Kullman and Öberg 2009;MoiseevandShiyatov2003;
Roush et al. 2007). Stockdale et al. (2015)introduced
a method of quantifying land cover change using the
WSL-MT and repeat photography from the Mountain
Legacy Project (MLP), a collection of over 120,000 his-
toricsurveyimagesoftheCanadianRockyMountains
(1888–1958; Trant et al. 2015). The study compared esti-
mated coordinates of tie points used in WSL-MT camera
calibration with the coordinates of points that had been
projected from oblique to orthogonal perspective using
the software; the mean error between points (error vec-
tor length) was 14.7 m, while mean displacement error
(i.e., dierence between the centroid of the set of all test
points in a given image) was 2.9 m (Stockdale et al. 2015).
The high spatial resolution of airborne lidar makes it a
promising candidate for validation of oblique photograph
analysis, but at the time of writing, only a single study is
known to have validated land cover data generated from
the WSL-MT using vegetation measurements from air-
borne lidar (Kolecka et al. 2015). In that study, lidar-
derived estimates of vegetation height and canopy cover
were compared to areas of forest succession mapped from
oblique photographs, with an overall accuracy of 95%.
Methods
Oblique photographs
Research was conducted over the West Castle Watershed
(WCW), Alberta, Canada. WCW is located in the head-
waters of the Oldman River on the eastern slopes of the
Rocky Mountains, with an area of ∼103 km2(watershed
dened as upstream of the University of Lethbridge West
Castle Field Station, located at 49.35°N, −114.41°W),
and an elevation range of ∼1400–2600 m above sea
level. Aside from a small ski resort and village near the
downstreamendofWCW(2.9%ofthewatershedarea),
and trail network in the valley bottoms, the forested
slopes demonstrate minimal anthropogenic disturbance,
andhavenotbeensubjecttomountainpinebeetle
infestation. Consequently, the WCW is an ideal study
area for assessing recent natural shifts in the treeline
ecotone in this part of the Canadian Rockies. Seven high-
resolution oblique photographs of WCW were selected
fromasetof46providedbytheMLP,whichmaximized
the spatial coverage of the valley while also providing
areasofoverlapforclassicationerroranalysis.All
photographs were collected in August 2006, and geoloca-
tions for photograph origins were provided by the MLP
(Table 1 ).
Rasterized canopy cover data were generated from
oblique photographs using WSL-MT, following the
Downloaded by [David McCaffrey] at 09:22 01 November 2017
506 D. R. McCAFFREY AND C. HOPKINSON
Tab le . Oblique photograph summary.
Photograph Number Elevation (m) Latitude Longitude Photograph Date
, °. −°. July ,
, °. −°. July ,
, °. −°. July ,
, °. −°. July ,
, °. −°. July ,
, °. −°. July ,
, °. −°. July ,
method of Stockdale et al. (2015): (i) co-registered aerial
imagery (1.5 m, SPOT 6, acquired July 31, 2014, © 2014
CNES, licensed by BlackBridge Geomatics; purchased by
Planet Labs Inc., 2015) and topographic data (1 m DEM,
lidar collection described below) were used to calibrate
camera parameters for each of the 7 images; (ii)foreach
image, an iterative process of control point selection was
used to reduce 21 potential control points to 6 control
points with the least angle error—discarded control point
became ‘test points’ used in error analysis; (iii)thecamera
parameters were exported to ArcGIS 10.3 (Esri 2014),
where the 1 m DEM was used to create a viewshed for
the image (Figure 1b); (iv) a 20-m shnet (i.e., 400 m2
grid cells) was overlaid and clipped to the viewshed,
with any grid cells having less than 75% coverage (i.e.,
<300 m2)beingomitted(Figure 1c); (v)theWo r l d to
Pixel function of the WSL software was used to project
shnet grid cells from UTM coordinates to pixel coor-
dinates, so that the 20 m ×20 m grid cells were draped
over the oblique image (Figure 1d). Once in an oblique
projection, grid cells were manually assigned into 1 of 4
canopy cover classes based on observed canopy openness
and texture (Figure 1e). The 4 canopy cover classes were:
(i)No Cover—grid cells devoid of vegetation; (ii)Low
Veget a t i on—grid cells appear vegetated, but context and
texture suggest shrubs or krummholz, not upright trees
>2m;(iii)Partial Canopy—trees are present, but ground
is visible in >50% of the grid cell; (iv)Full Canopy—trees
cover >50% of a grid cell. Finally, classied oblique
images were converted back to orthogonal view for
comparative analysis (Figure 1f).
Visual estimates of canopy cover were completed by
a single trained interpreter. While it is well known that
manual photograph interpretation is subject to error
introduced by misregistration and interpreter bias, which
precludes single interpreter analysis, recent studies of
fractional cover assessment from aerial imagery have
established a multi-interpreter agreement using a second
interpreter, working on a subset (5%–10%) of the data
(Nowak and Greeneld 2012;Ucaretal.2016). We ran-
domly selected 5% of the observed grid cells, and used
a second interpreter to classify canopy cover using the
same 4-class criteria.
Error analysis
Error vector length and displacement errors were cal-
culated for each of the 7 photographs in the 2006 MLP
dataset, using the discarded test points from the camera
calibrations. Error vector length is the 3-D distance
between each test point and its projected counterpart,
and displacement error is the 3-D distance between the
centroid of the set of test points and the set of projected
points. Additionally, landscape level measurements of
error vector length and displacement were calculated
using all available test points (n=64), without stratifying
by image.
Using the Alberta Vegetation Inventory (Alberta
Environment and Parks 2012), vector data for known
areas of disturbance (e.g., former cut blocks, oil well pads,
roads, ski resort, and a small burn area) were aggregated
into a single layer, and any grid cells intersecting with this
disturbance layer were omitted. Grid cells that extended
above ridgelines or the horizon were omitted from analy-
sis, leaving a nal observation area of 38.5 km2, or 37.7%
of WCW.
To understand factors that contribute to error vector
length, Stockdale et al. (2015) modeled vector error
length as a function of distance from the camera origin,
and angle of viewing incidence. Angle of viewing inci-
dence is dened as the angle between a ray connecting
the camera origin to the control point, and a 10-m line
segment xed to the ground along the sight line of the
ray, with the control point as a midpoint. A general linear
model was executed in SPSS 24 (IBM 2016), using the
formula:
Error Vector Length (m)
=Intercept +β∗I+β∗D+β∗A,[1]
where Iis image number (used as a random factor),
Dis distance to camera, Ais the angle of viewing
incidence, and βis the model coecient for each
parameter.
Additionally, an ‘observation-parameter routine’ was
constructed, which calculates the distance to the camera
and viewing incidence for any set of points on the DEM,
in ArcGIS 10.3 (ESRI 2014), allowing the modeled error
Downloaded by [David McCaffrey] at 09:22 01 November 2017
CANADIAN JOURNAL OF REMOTE SENSING 507
Figure . The oblique photograph to raster workflow, as described by Stockdale et al. (): (a) an oblique image of Mount Haig, in the
West Castle Watershed, copyright Mountain Legacy Project (); (b) orthogonal view shed of oblique Mount Haig image, using camera
calibration from WSL Monoplotting Tool; (c) a -m fishnet applied to the Mount Haig view shed; (d) fish net grid projected back to
oblique view using WSL World to Pixel function; (e) oblique image with canopy cover classification; and (f) orthogonal image with canopy
cover classification.
vector length to be applied at any observation grid cell in
the watershed.
MLP photographs with overlapping extents were
intentionally selected to test correspondence between
separate observations. Of the total observed area,
10.5 km2(27.2%) had multiple observations, which
enabled an assessment of the manual classication con-
sistency from dierent vantage points (Figure 2a,Ta b le 2).
Accuracy of canopy cover classications was assessed by
determining whether the minimum and maximum clas-
sication values for each grid cell agreed or disagreed. In
cases of disagreement between observations, the canopy
cover classication from the image with the lowest error
vector length, as modeled by the observation-parameter
routine, were selected.
Airborne lidar
An airborne lidar survey of WCW was own on Octo-
ber 18, 2014 with a Leica ALS70 and a minimal point
spacingof3pts/m
2at nominal altitude (1,300 m above
the mean ground surface elevation). Lidar point cloud
Downloaded by [David McCaffrey] at 09:22 01 November 2017
508 D. R. McCAFFREY AND C. HOPKINSON
Figure . Multiple observations: (a) observation areas were viewed by either , , or different oblique images and (b) areas with multiple
observation either agreed in all or observations, or disagreed.
processing and fractional cover analysis followed the
canopy-to-total returns method described in Hopkinson
and Chasmer (2009).Datawereclassiedintogroundand
non-ground using Ter ra s can (Terrasolid, Finland) and a
1-m DEM was interpolated from the ground-classied
points and aggregated using mean values to match the
20-m canopy cover classication grid cells. Fractional
cover was calculated using the LAStools suite (Isenburg
2013); data were height normalized, and lascanopy was
applied with a height threshold of 2.0 m and a step of
20 m, again matching canopy cover classication grid
(Figure 3d).
Fractional cover, as a metric, treats canopy cover as
acontinuousvariable(i.e.,apercentage).Tocompare
the agreement between factional cover and manually
classied canopy cover, which has discrete, ordinal
classes, we needed to discretize the measurements of
fractional cover into 4 classes. No logical break points
existed a priori, so we used 3 common, unsupervised
discretization methods: equal frequency, equal range,
and k-means. We tested agreement between 4-class
manual interpretation and ordinal, binned classes of
fractional cover using Cohen’s K, weighted linearly
(Cohen 1968).
Tab le . Agreement of multiple observations.
Number of Observations Area (km) Percent of Observed Area
..
. .
..
Multiple Observations
Agree . .
Disagree . .
As the results may have been inuenced by expected
areas of homogenous full canopy cover at lower elevations
and no cover at higher elevations, each analysis was run
twice; once on all observed areas, and once restricted
to elevations between 1800 m–2300 m, where ATE is
expectedtocausemorevariabilityincanopycover.
Results
Overall correspondence between the primary interpreter
and the secondary interpreter was 88.4%, and weighted
Cohen’s K showed substantial agreement between inter-
pretations (K =0.79, p<0.001). When elevations were
restricted to the more variable ATE region, correspon-
dence dropped to 74.1%, with weighted K again showing
substantial agreement (K =0.73, p<0.001).
Mean error vector length between test points ranged
from 2.0 m to 63.8 m per image, with a mean of 23.9 m
among the 7 images (Tab l e 3). Displacement ranged
from 0.8 m to 39.5 m, with a mean of 14.4 among the 7
images. In the landscape level analysis of 64 aggregated
points, mean error vector length was 21.7 m and mean
displacement was 7.0 m.
A general linear model of error vector length was
used,with2adjustmentstothedatasetof64testpoints:
(i) 2 outlier points were identied and removed—these
outliers had error vector length values of 177.7 m and
192.1 m (the next highest value was 78.0 m), and occurred
in dierent images (photos 6 and 3, respectively); (ii)a
Shapiro-Wilk test demonstrated the data were not nor-
mally distributed (W=541, p<0.001)—the truncated
sample of 62 values responded to a common log trans-
formation, and a repeat of the Shaprio-Wilk test failed to
reject normality (W=9.65, p=0.076). We proceeded
Downloaded by [David McCaffrey] at 09:22 01 November 2017
CANADIAN JOURNAL OF REMOTE SENSING 509
Figure . West Castle Watershed area of interest: (a) minimum error vector length, as modeled by the GLM; (b) number of photograph
with lowest modeled error, used for canopy cover classification; (c) four classes of manual canopy cover classification; and (d) lidar-derived
fractional cover.
Tab le . Error analysis for WSL-MT camera calibration, error vector length, and modeled error.
Photo
No. of
Registration
Points
Mean Angle
Error (°)
No. of Test
Points
Mean Error Vector
Length (±SE)
Mean
Displacement
Error (m)
Mean Distance From Camera
(m) (range)
Mean Angle of Viewing
Incidence (°)(range)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
. . (.) . ,. (,.–,.) . (.–.)
Mean . . . (.) . ,. .
Landscape . (.) .
. (.) .
Model Error Vector Length , . (<.) .–.
Downloaded by [David McCaffrey] at 09:22 01 November 2017
510 D. R. McCAFFREY AND C. HOPKINSON
Figure . Lidar-derived fractional cover values were discretized into four ordinal groups using three methods (i.e., equal frequency, equal
range, and k-means); the boxplot depicts the IQR for each of these three groups (in blue shades) and compares their fractional cover
distributions against those generated from manual classifications (in red).
Tab le . Agreement between manual classifications and discretized fractional cover.
All ATE
Discretization
Method
Accuracy
(%)
Weighted
K
Standard
Error
Lower
% C.I.
Upper
% C.I. P Value
Accuracy
(%)
Weighted
K
Standard
Error
Lower
% C.I.
Upper
% C.I. P Value
Equal frequency . . . . . <. . . . . . <.
Equal range . . . . . <. . . . . . <.
K-means . . . . . <. . . . . . <.
Manual breaks . . . . . <. . . . . . <.
with the general linear model, which yielded the following
equation:
log Error Vector Length (m)
=0.992 +β∗I+0.000077∗D+0.009∗A.[2]
All factors were signicant at α=0.05, except for
distance to camera (D;f=1.278, p=0.263). The
observation-parameter routine was used to measure the
distance to camera and angle of viewing incidence for
each grid cell in the analysis produced a surface map
of the modeled error vector length (Figure 3a). Mean
error vector length for the model was 14.6 m, and ranged
between0.9mand75.7m(Ta b le 3); these values were
comparabletothemeanandrangeofthedatasetusedto
develop the model.
Theanalysisofmultipleobservationsshowedthat
no area in the watershed was observed in more than 3
dierent images (Figure 2a). Comparison of multiple
observation areas showed that 78.4% of the grid cells
agreed between observations, while at least 1 classi-
cation value disagreed in 21.6% of cases (Figure 2b,
Table 2 ). Agreement between manual cover classica-
tions and fractional cover classes was generally consistent
across discretization methods (Figure 4); 4-class accu-
racy ranged from 43.3%–47.0%, and weighted K values
showing moderate agreement (K =0.42–0.43, p<0.001;
Table 4 ). In the ATE, 4-class accuracy was only slightly
reduced from the full watershed analysis, ranging from
39.4%–46.0%, with weighted K values showing fair
agreement (K =0.35–0.36, p<0.001; Tab le 4).
Discussion and conclusion
The analyses indicated that fractional cover across the
ATE can be discriminated from oblique imagery at a res-
olutionof20m.Meanerrorvectorlengthandmeandis-
placement error were both substantially higher than error
values calculated in Stockdale et al. (2015); our study cal-
culated a mean error vector length of 23.9 m and a land-
scape level displacement error of 7.0 m, while Stockdale
et al. (2015)found14.4mand2.9merrors,respectively.
Followingtheremovalof2outliers,meanerrorvector
length and displacement error were more comparable to
the previous study, at 16.5 m and 5.9 m. Several possi-
bilities exist for the larger error observed in the present
study. Stockdale had a larger sample size of test points
Downloaded by [David McCaffrey] at 09:22 01 November 2017
CANADIAN JOURNAL OF REMOTE SENSING 511
(n=121), which resulted from fewer omissions in test
point projection from WSL-MT. It should be noted that
these omissions may reect higher error vector lengths
than those seen in processed test points, thus making
actual error vector length in the present study higher than
what is reported. Another potential cause of the discrep-
ancy between error estimates is that aerial imagery used
for control point selection in the previous study was at a
resolution of 0.5 m, 3 ×greater resolution than the 1.5 m
SPOT 6 data used in the present study. Higher resolution
aerial imagery could have improved control point place-
ment, reduced angle error, and contributed to a lower
mean error vector length. Regardless of the discrepancy,
in the present study mean error vector length was com-
parable to canopy cover grid cell resolution and mean
displacement error remained substantially below this res-
olution, demonstrating that fractional cover can be ras-
terized from oblique photographs at a resolution of 20 m.
This spatial resolution is an important threshold to reach
for the application of the method to ATE observations, as
it corresponds to the area of most forest mensuration plots
(400 m2), and permits cataloguing of changes in the ATE
given the expected rate of advance over the century scale
observation period permitted by repeat photography.
Comparison of grid cells with multiple observations
showed that canopy cover classication disagreed in
21.6% of cases. Disagreement cases were resolved using
modeled error vector length, which produced values
comparable to those in the observed error vector length;
observed mean error vector length for the 62 test points
used for model development was 16.5 m, with a range
of 1.9 m–78.0 m, while modeled values had a mean
of 14.6 m and a range of 0.9 m–75.7 m (Table 3 ). The
previous study found that image number (I), distance
to camera (D), and angle of viewing incidence (A)all
contributed signicantly to the model (i.e., p<0.05), but
inourmodelthefactorofdistancetocamera(D)was
not signicant. These results demonstrate that models
of error vector length can be useful in resolving cases of
disagreement in oblique photograph classication, and
thatangleofviewingincidencecanbeconsideredthe
primary factor aecting displacement error.
Manual classications of fractional cover from the
oblique photographs were, in part, validated using lidar-
derived fractional cover. Fractional cover distributions
from manual classications had increasing median val-
ues with each ordinal class, in both the full watershed
and ATE restricted analyses. While accuracy between
the discretized fractional cover classes and the manual
classes was generally low, peaking at 47.0%, it is worth
considering this may have resulted from the fact that the
discretization methods used were theoretical, and did
not necessarily reect real-world bins of fractional cover.
For example, if we arrange lidar-derived fractional cover
values in order, and partition them into bins with equal
frequency ratios as those seen in the manual classica-
tions, we reach accuracy upwards of 64.7% (see Table 4 ,
Manual breaks). While this method is not a defensible
validation, as it assumes that the manual classications
aretruth,itdoesdemonstratethattheresultsofthis
analysis rely on partition points that are in some ways
arbitrary.
With an abundance of repeat oblique photography
in the Canadian Rocky Mountains (Trant et al. 2015),
future research, which combines quantitative analysis of
oblique photography with lidar-derived forest structure
data, can increase the spatial resolution and temporal
extent of ATE monitoring. This is important as we con-
tinue to explore the impacts of climatic change, natural
and anthropogenic disturbances on forested mountain
ecosystems, and downstream water resources.
Acknowledgments
Wewouldliketothanktheeditorandanonymousreviewersfor
their helpful comments. We also wish to thank Dr. Eric Higgs
andDr.MarySanseverinooftheMountainLegacyProject
for providing oblique photographs, and Dr. Claudio Bozzini
for instruction on the WSL monoplotting tool. Additionally,
we gratefully acknowledge Josh Montgomery for acting as the
second photograph interpreter.
Funding
ThisresearchwasmadepossiblebygrantsfromAlbertaInno-
vates, Environment and Energy Solutions, Water Innovation
Program (Grant #E323726), NSERC Discovery (Grant #2017-
04362), and the Canadian Foundation for Innovation (Grant
#32436).
References
Alberta Environment and Parks. 2012. Alberta Vegetation
Inventory. Edmonton, Canada: Alberta Environment
and Parks.
Bekker, M.F. 2005. “Positive feedback between tree establish-
ment and patterns of subalpine forest.” Arctic, Antarctic, and
Alpine Research, Vol. 37(No. 1): pp. 97–107.
Bozzini, C., Conedera, M., and Krebs, P. 2012. “A new mono-
plotting tool to extract georeferenced vector data and
orthorectiedrasterdatafromobliquenon-metricpho-
tographs.” International Journal of Heritage in the Digital
Era, Vol. 1(No. 3): pp. 499–518.
Butler, D.R., and DeChano, L.M. 2001. “Environmental change
in Glacier National Park, Montana: An assessment through
repeat photography from re outlooks.” Physical Geography,
Vol. 22(No. 5): pp. 291–304.
Case, B.S., and Duncan, R.P. 2014. “A novel framework for dis-
entangling the scale-dependent inuences of abiotic factors
Downloaded by [David McCaffrey] at 09:22 01 November 2017
512 D. R. McCAFFREY AND C. HOPKINSON
on alpine treeline position.” Ecography,Vol.37(No.9):
pp. 838–851.
Cohen, J. 1968. “Weighted kappa: Nominal scale agreement pro-
vision for scaled disagreement or partial credit.” Psychologi-
cal Bulletin, Vol. 70(No. 4): pp. 213.
Coops, N.C., Morsdorf, F., Schaepman, M.E., and Zimmer-
mann, N.E. 2013. “Characterization of an alpine tree line
using airborne LiDAR data and physiological modeling.”
Global Change Biology, Vol. 19(No. 12): pp. 3808–3821.
Danby, R.K. 2011. “Monitoring forest-tundra ecotones at multi-
ple scales.” Geography Compass, Vol. 5(No. 9): pp. 623–640.
ESRI. 2014. ArcGIS 10.3 for Desktop.Redlands,CA:Environ-
mental Systems Research Institute.
Hagedorn, F., Shiyatov, S.G., Mazepa, V.S., Devi, N.M., Grigorév,
A.A., Bartysh, A.A., Fomin, V.V., Kapralov, D.S., Terentév,
M., Bugman, H., Rigling, A., and Moiseev, P.A. 2014. “Tree-
line advances along the Urals mountain range—Driven
by improved winter conditions?” Global Change Biology,
Vol. 20(No. 11): pp. 3530–3543.
Holtmeier, F.K., and Broll, G. 2005. “Sensitivity and response
of northern hemisphere altitudinal and polar treelines to
environmental change at landscape and local scales.” Global
Ecology and Biogeography, Vol. 14(No. 5): pp. 395–410.
Hopkinson, C., and Chasmer, L. 2009. “Testing LiDAR models
of fractional cover across multiple forest ecozones.” Remote
Sensing of Environment, Vol. 113(No. 1): pp. 275–288.
IBM. 2016. SPSS Statistics for Windows. Armonk, NY: IBM
Corp.
Isenburg, M. 2013. LAStools-ecient tools for LiDAR process-
ing. Version 161114. Open-sourced software. Gilching, Ger-
many: rapidlasso.
Klasner, F.L., and Fagre, D.B. 2002. “A half century of change in
alpine treeline patterns at Glacier National Park, Montana,
U.S . A .” Arctic, Antarctic, and Alpine Research,Vol.34(No.1):
pp. 49–56.
Kolecka, N., Kozak, J., Kaim, D., Dobosz, M., Ginzler, C., and
Psomas, A. 2015. “Mapping secondary forest succession
on abandoned agricultural land with LiDAR point clouds
and terrestrial photography.” Remote Sensing,Vol.7(No.7):
pp. 8300–8322.
Körner, C. 2012. Alpine Treelines: Functional Ecology of the
Global High Elevation Tree Limits.NewYork,NY:Springer
Science & Business Media.
Kullman, L., and Öberg, L. 2009. “Post-Little Ice Age tree
line rise and climate warming in the Swedish Scandes: A
landscape ecological perspective.” Journal of Ecology,
Vol. 97(No. 3): pp. 415–429.
Liang, E., Dawadi, B., Pederson, N., and Eckstein, D. 2014.
“Is the growth of birch at the upper timberline in the
Himalayas limited by moisture or by temperature?” Ecology,
Vol. 95(No. 9): pp. 2453–2465.
Luckman, B., and Kavanagh, T. 2000. “Impact of climate uc-
tuations on mountain environments in the Canadian Rock-
ies.” Ambio: A Journal of the Human Environment,Vol.29
(No. 7): pp. 371–380.
Moiseev, P.A., and Shiyatov, S.G. 2003. “Vegetation dynamics
at the tree-line ecotone in the Ural Highlands, Russia.” In
Alpine Biodiversity in Europe,editedbyL.Nagy,G.Grab-
herr, C. Körner, and D.B.A. Thompson, pp. 423–435. Berlin,
Germany: Springer Science and Business Media.
Nowak, D.J., and Greeneld, E.J. 2012. “Tree and impervious
cover change in U.S. cities.” Urban Forestry & Urban Green-
ing, Vol. 11(No. 1): pp. 21–30.
Roush, W., Munroe, J.S., and Fagre, D.B. 2007. “Development
of a spatial analysis metho using ground-based repeat pho-
tography to detect changes in the alpine treeline ecotone,
Glacier National Park, Montana, USA.” Arctic, Antarctic,
and Alpine Research, Vol. 39(No. 2): pp. 297–308.
Stockdale, C.A., Bozzini, C., Macdonald, S.E., and Higgs,
E. 2015. “Extracting ecological information from oblique
angleterrestriallandscapephotographs:Performanceeval-
uation of the WSL Monoplotting Tool.” Applied Geography,
Vol. 63: pp. 315–325.
Tranquillini, W. 1979. Physiological Ecology of the Alpine Timber-
line. Tree Existence at High Altitudes with Special Reference
to the European Alps. Berlin, Germany: Springer-Verlag.
Trant, A.J., Starzomski, B.M., and Higgs, E. 2015. “A publically
available database for studying ecological change in moun-
tain ecosystems.” Frontiers in Ecology and the Environment,
Vol. 13(No. 4): pp. 187–187.
Ucar, Z., Bettinger, P., Merry, K., Siry, J., Bowker, J.M.,
and Akbulut, R. 2016. “A comparison of two sampling
approaches for assessing the urban forest canopy cover
from aerial photography.” Urban Forestry & Urban Greening,
Vol. 16: pp. 221–230.
Weiss, D.J., Malanson, G.P., and Walsh, S.J. 2015. “Multi-
scale relationships between alpine treeline elevation and
hypothesized environmental controls in the western United
States.” Annals of the Association of American Geographers,
Vol. 105(No. 3): pp. 437–453.
Downloaded by [David McCaffrey] at 09:22 01 November 2017