positives. Each of the three
NDVI
change detection methods
could potentially identify those false detections from
NLCD
data. For the minimum-value method, if the
NDVI
-derived
change year located outside of the period of 2001–2011, we la-
bel the pixel as no-change. For the break-point method, a nega-
tive value of µ1–µ2 signals an increase of
NDVI
value in the
time-series so it is reasonable to label the corresponding pixel
as no-change. For simple-threshold identification, the no-
change pixel can be determined based on a predefined thresh-
old (e.g., 0.6). For a given
NDVI
time-series, if all
NDVI
values
within the 2001–2011 window are higher than the threshold
value of 0.6, the pixel could be labeled as no-change.
Accuracy Assessment
We conducted detailed accuracy assessments for two change
pixel groups: (1) an urbanization group that includes all
pixels changed from nonurban land cover classes to medium-
high intensity urban, and (2) an urban intensification group
that includes pixels changed from open space or low intensity
urban to medium-high intensity urban. We systematically
selected a total of 400 pixels for accuracy assessment: 200
for the urbanization group and 200 for the urban intensifica-
tion group. Within the urban intensification group, 100 pixels
were selected for open space to medium-high intensity urban
change and low intensity to medium-high intensity urban
change, respectively.
Each selected change pixel was pinpointed on Google
Earth and the actual urban change year was recorded by vi-
sual interpretation of Google Earth’s High-Resolution Imagery
Archive. The conversion of forest/agriculture cover to urban
cover can be a multiyear process from the beginning of clear-
cutting to the completion of construction. We used the begin-
ning of clear-cutting as the “ground truth” of urban change.
In some cases, it was difficult to detect the starting point of
clear-cutting due to limited availability of historical satellite
images. For example, an ongoing construction was observed
in the March of 2005 Google imagery and the previous July
2003 Google image showing forest/agricultural cover, the mid-
dle temporal point (year 2004) was then selected as the urban
change year. For pixels without sufficient high-resolution
imagery as reference (i.e., imaging gap > two years), we sim-
ply excluded them from the accuracy assessment because we
could not define a “ground truth” or o
year. About 29 pixels were removed be
gibility for the accuracy assessment, a
were selected by simple random sampling. The same method
was applied to the urban intensification pixel group for deter-
mining change years. The only difference is that there might
not be a clear-cutting event so the observed urban intensifica-
tion year should be the year when construction started.
The urban change years estimated from the
NDVI
time
series analyses were compared with the Google Earth-derived
data, pixel-by-pixel. Due to image availability issues, the
Google Earth-derived urban change year could have a one-
year deviation from the actual urban change year, therefore,
one-year difference between the
NDVI
-estimated change year
and the Google Earth-derived change year was considered as a
“correct” change detection. Overall accuracies were comput-
ed to compare three change detection algorithms using such a
one-year deviation assessment. Additional two-year deviation
assessments were conducted for thoroughness.
Results
Accuracy Assessment for Urbanization Pixels
Table 1 shows the error matrix of accuracy assessment for
change and no-change pixels using Google Earth as reference.
Note all 200 pixels were previously identified as change pixels
(from other land cover classes to medium-high intensity urban)
by comparing
NLCD
2001 and 2011. Among the 200 selected
pixels, 196 pixels were visually interpreted as change pixels
using high resolution Google Earth imagery, suggesting very
high accuracy of
NLCD
in determining other land cover classes
(e.g., forest and agricultural lands) to medium-high intensity
urban change. Using 1998–2014
NDVI
time-series data as input,
all three time-series methods performed well on the separa-
tion of change and no-change pixels, and the minimum-value
method had the highest overall accuracy of 98.5% (kappa =
0.56). Simple-threshold identification performed worst by
falsely identifying no-change pixels, leading to a high commis-
sion error of 71.4%. The use of longer time-series
NDVI
(1988–
2017) resulted in almost identical overall accuracy (difference
< 1%) so the detailed error matrix is not presented here.
Table 1 summarizes general accuracy statistics for change
and no-change pixels only. Figure 4 shows a scatter plot com-
paring the
NDVI
-derived and the Google Earth-derived change
year, pixel-by-pixel. For simplicity, we used results from the
as an example to demonstrate the com-
randomly selected points, ten were inter-
pixels either from Google Earth or from the
algorithm. The remaining points were plotted with different
Table 1. Error matrices of change and no-change for urbanization pixels using reference data derived from Google Earth for time
series 1998–2014.
Parameter
Reference from Google Earth
No-Change
Change
Total
%Correct
%Commission
Minimum-value method
No change
2
1
3
66.7
32.3
Change
2
195
197
99.0
1.0
Total
4
196
200
98.5
(n = 200)
%Correct
50.0
99.5
%Omission
50.0
0.5
kappa = 0.56
Break-point method
No change
2
6
8
25.0
75.0
Change
2
190
192
99.0
1.0
Total
4
196
200
96.0
(n = 200)
%Correct
50.0
97.0
%Omission
50.0
3.0
kappa = 0.32
Simple-threshold (t = 0.6)
No change
4
14
18
22.2
77.8
Change
0
182
182
100.0
0
Total
4
196
200
92.2
(n = 200)
%Correct
100.0
92.6
%Omission
0
7.4
kappa = 0.34
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2019
719
