D
ata
and
M
ethods
Change detection and labeling were both based on monthly
continental composites of MODIS reflectance observations
acquired for 2005 and 2010. The composites were produced by
theCCRS followingmethodologies described inKhlopenkov and
Trishchenko (2008), Luo et al. (2008), and Latifovic et al. (2012).
Identifying land cover change required a two-step protocol: (1)
detecting areas of potential change and (2) assigning a land
cover class to those areas, similar to that described in Latifovic
and Pouliot (2005), Xian et al. (2009), Fry et al. (2011), Pouliot
et al. (2012), and Pouliot et al. (2014). Unlike other approaches
such as the global MODIS land cover maps (MCD12Q1; Friedl
et al. 2010), this two-step approach assures high consistency
of mapping in a land cover time series, because areas where
change has not occurred remain unaffected and change areas
can be directly compared across time.
Prior to change detection, the initial North America land
cover map of 2005 (LC2005) was modified, and the improved
version (LC2005V2) was used as baseline for generating the
2010 land cover product. This was necessary because the
updating procedure required an accurate baseline as only
areas of potential change will be altered and mislabeled
patches or false detection would have negatively affected
change results in their spatial context.
The actual implementation of the updating procedure varied
among the countries because of existing prior approaches
and additional custom needs of each participating country.
Country-specific procedures did not notably affect the overall
map consistency. For the first step of change detection, the
binary mask of change and no change for Canadian forests
was derived using a regression tree procedure for abrupt
disturbances (Guindon et al. 2011) and a multifaceted trend-
based approach for detection of more gradual and subtle
changes (Pouliot et al. 2014). The United States applied a
complex change vector approach mainly based on the NDVI
and red and near infrared bands (Jin et al. 2013). For Mexico,
a data-driven process that uses difference images of all
bands, the NDVI, and edge detection filters from all months
was developed (Colditz et al. 2014).
For the second step of change labeling, each country
also conducted its own specific classification approach yet
followed the same general procedure: training data were
sampled from surrounding unchanged areas and the spectral
data set of 2010 served as features to generate a decision tree
that was subsequently applied to the potential change areas.
Additional post-processing steps involved evidence-based
spatial and contextual reasoning for Canada, localized GIS
modeling in the U.S., and a transition matrix to only allow
logical changes in Mexico.
R
esults
Results are presented in the following four sections: (1)
improvements to the 2005 base map, (2) country-specific
changes between this improved map and the updated
land cover for year 2010, including preliminary results for
accuracy assessment, (3) class-specific changes, and (4) site-
specific analysis of detected land cover changes. The results
presented here were derived for classifications at level 2 (19
classes) and a minimum mapping unit of 25ha. The state of
Hawaii was excluded from this analysis.
Difference Between the Original and Improved Land
Cover Maps of 2005
A total area of 1,077,701km
2
(5.06% of the North American
continent) was modified in the improved base line map
(LC2005V2). For Canada, further quality assessment and
input from users suggested modifications in some areas (in
total 638,963km
2
, 6.45% of the total area of Canada). The
major changes included reduction of treed wetland along
the western and eastern coasts, correction of water/shadow
confusion in mountain areas, enhanced shrub classification,
and some local improvements to wetland mapping.
Improvements in the United States entirely focused on the
state of Alaska, which involved a complete reprocessing due
to a spatial mis-registration of the initial 2005 land cover map
for this state (429,916km
2
, 4.55% of the total area of the U.S.),
while the land cover map of the lower 48 states was deemed
sufficient for map updating. In Mexico, improvements focused
on two land cover classes (8,821km
2
, 0.45% of the total area
of Mexico): urban and built-up areas were underestimated or
omitted due to a low spectral contrast to sparse vegetation
in semi-arid northern Mexico and class Water, which was
generally underestimated in the original map.
Identifying land cover change required a
two-step protocol: (1) detecting areas of
potential change and (2) assigning a land
cover class to those areas.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2014
919
