nonparametric machine learning classifier (the random forest)
rather than a parametric maximum likelihood classifier and
stable site identification procedures that are adapted based on
the prevalence of each individual class rather than constant
across all classes. Com-
pared to other common
classification methods,
the random forest classi-
fier is relatively accurate
and robust to noise (Brei-
man, 2001; Dannenberg
et al
., 2016; Gislason
et al
., 2006; Rodriguez-
Galiano
et al
., 2012),
which is an important
property for
AASG
since
it should mitigate inac-
curacies in the reference
classification, stable site
identification proce-
dures, and the predictor
variables. Here, we used
the randomForest pack-
age in the R statistical
computing environment
with an ensemble of 250
trees and with prior probabilities set based on the prevalence
of each class in the 2011
NLCD
reference classification.
Like Gray and Song (2013), we identified a training set of
stable sites from those pixels that were nearest to the mean of
the image difference histogram (I
2
−I
1
). However, rather than
using a constant standard deviation threshold for all classes,
we adapted the proportion of pixels identified as stable based
on the prevalence of each class in order to balance the com-
peting goals of achieving a large training set while minimiz-
ing the inclusion of nonstable sites in the training set. Thus,
more stringent thresholds (<<10% of pixels) were used for
classes with high areal cover to minimize the likelihood of
misidentified stable sites, while more liberal thresholds (10-
50% of pixels) were used for classes with low areal cover to
achieve an adequate training set. Based on previous experi-
ments (Dannenberg
et al
., 2016) and experiments conducted
for the current work, we set a large upper limit on the maxi-
mum number of pixels that could be identified as stable for
any given class (up to 500,000 per land cover class per scene,
depending on the overall prevalence of the class in the refer-
ence classification).
Multi-season Landsat Imagery and Preprocessing
Our updated classifications were developed from Landsat
Collection 1 surface reflectance (Masek
et al
., 2006) and
brightness temperature from thermal band 6. We used the
2011
NLCD
as the reference classification, C
1
, with image I
1
selected as the Landsat-5
TM
image with the minimum cloud
cover during May-September within the years 2010-2012.
However, we masked out any pixels classified as “Developed,
Open Space” (class 21) in the
NLCD
since this class includes
a disparate array of pervious and impervious surfaces, most
commonly suburban lawns and dwellings as well as parks,
golf courses, and other urban vegetation (Multi-Resolution
Land Characteristics Consortium, 2017). Since spectral sig-
natures of “Developed, Open Space” are neither consistent
nor distinct from other classes, additional ancillary data is re-
quired to identify this class, and therefore, following previous
studies (Sexton
et al
., 2013), we chose to exclude “Developed,
Open Space” from our classifications.
For classifications of nominal years 1986, 1991, 1996, and
2006, we used primarily Landsat-5
TM
imagery supplemented
with several Landsat-4
TM
images for the 1986 and 1991 clas-
sifications. The classification of nominal year 2001 was based
on approximately 60% Landsat-5
TM
images and roughly 40%
Figure 2. Classification workflow. Inputs to
AASG
include
a reference classification (C
1
), a reference image (I
1
)
corresponding to C
1
, and a set of images to be classified
(I
2
). Here, I
1
and I
2
were drawn from a 3-year window
surrounding the nominal classification year and were
converted to tasseled cap brightness, greenness and wetness
components. For purposes of stable site identification,
we used growing season brightness images. For purposes
of classification, I
2
consisted of a set of tasseled cap
components from three multi-season images (unless cloud-
and snow-free images were unavailable for a given season)
plus ancillary topographic metrics derived from a digital
elevation model from
SRTM
.
Table 1. Agreement between the 2001/2006
AASG
classifications and the withheld 2001/2006
NLCD
classifications, excluding pixels classified as Developed, Open Space in the
NLCD
.
OA*
: overall agreement;
f
c
: fraction of total area mapped as each class;
PA*
: producer’s agreement;
UA*
: user’s agreement.
2001 (OA* = 75.7%)
2006 (OA* = 72.7%)
Class
f
c
NLCD f
c
AASG PA* UA* f
c
NLCD f
c
AASG PA* UA*
11: Water
4.9% 4.9% 92.8% 93.2% 5.0% 5.0% 93.0% 92.0%
22: Developed, Low Intensity
2.3% 2.6% 63.0% 55.8% 2.3% 2.6% 61.9% 55.1%
23: Developed, Medium Intensity 0.7% 0.8% 58.1% 54.7% 0.8% 0.9% 56.4% 54.3%
24: Developed, High Intensity
0.3% 0.3% 72.9% 65.7% 0.3% 0.4% 71.2% 64.6%
31: Barren
0.3% 0.3% 43.8% 57.0% 0.4% 0.3% 39.9% 57.7%
41: Deciduous Forest
23.1% 22.5% 82.1% 84.3% 22.9% 22.6% 80.6% 81.6%
42: Evergreen Forest
10.9% 11.0% 74.8% 74.5% 10.6% 11.1% 71.2% 67.8%
43: Mixed Forest
3.2% 2.1% 32.5% 49.4% 3.1% 1.8% 25.9% 43.8%
52: Shrub/Scrub
6.0% 5.9% 56.4% 56.9% 6.2% 6.1% 47.2% 48.0%
71: Grassland/Herbaceous
8.8% 7.7% 64.5% 73.0% 9.2% 7.8% 59.4% 70.2%
81: Hay/Pasture
14.4% 16.3% 80.4% 71.3% 14.2% 16.3% 78.1% 68.1%
82: Cultivated Crops
14.6% 15.3% 86.1% 82.2% 14.6% 14.9% 83.2% 81.3%
90: Woody Wetland
8.4% 8.4% 72.0% 71.6% 8.4% 8.3% 70.4% 71.0%
95: Emergent Herbaceous Wetland 2.1% 2.0% 74.1% 78.0% 2.1% 2.0% 71.2% 75.4%
Table 2. Agreement between the simplified
NLCD
and
AASG
classifications.
OA*
: overall agreement;
PA*
: producer’s
agreement;
UA*
: user’s agreement.
2001
(OA* = 87.0%)
2006
(OA* = 85.3%)
Class
PA* UA* PA* UA*
Water (11)
92.8% 93.2% 93.0% 92.0%
Developed (22, 23, 24)
78.3% 70.3% 78.6% 71.4%
Barren (31)
43.8% 57.0% 39.9% 57.7%
Woody Vegetation
(41, 42, 43, 52, 90)
90.7% 93.8% 89.9% 92.0%
Herbaceous Vegetation (71, 81)
81.1% 78.3% 77.9% 75.8%
Cultivated Crops (82)
86.1% 82.2% 83.2% 81.3%
Emergent Herbaceous Wetland (95) 74.1% 78.0% 71.2% 75.4%
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2018
561
