Professional Documents
Culture Documents
change include Avalanche, Fire, Mass Movement, Progressive Defoliation (insects and disease),
Riparian, and Tree Toppling (wind throw and root rot).
The remaining change patches within the study area were assigned using Random forest
classification model in R (statistical software). R assigned each of the 412 change polygons to an
agent of change based upon similar attribute data and an out-of-bag error was assigned to the
complete dataset to determine accuracy. The dataset was further analyzed to determine the
coverage of each agent of change per year and within the entire study area.
RESULTS
The assignment of agents of change to the sample dataset using visual interpretation
resulted in an out-of-bag error of 41.82% for the entire NOCA study area, where change patch
polygons were classified to likely agents of change with approximately 58.18% accuracy based
upon attribute data correlation (figure 1). According to the Random Forest classification, all
change patch polygons assigned to the Avalanche agent of change were incorrectly assigned,
where the four polygons would be more accurately assigned to change agents of Mass Movement
(2), Progressive Defoliation (1), and Riparian (1). Progressive Defoliation was assigned with the
least amount of error with only 1 of 22 polygons misclassified as Tree Toppling.
The total 55 sample patches with detected changes between 2007 and 2009 covered
193.68 hectares (table 1). Sample change patch polygons were assigned to agents of change
using image interpretation. A majority of sample change patches were assigned to the agent of
change of Fire (table 2), where approximately 36% of the sample area was assigned to Fire agent
of change. Conversely, the least amount of change was assigned to Riparian agent of change at
4% of overall change patches. Also, a majority of the sample data had detected change in 2008
and the least amount of detected change occurred in 2009 (figure 2). These sample assignments
were applied to the Random Forest Classification to create a predicted dataset for the full
Landtrendr dataset of the NOCA study area.
Using the entire dataset of change patches in the NOCA study area, most change was
detected in the year 2007, followed by 2009, and the least in 2008 at approximately 15% (table
3). The Random Forest classification for the entire NOCA Study area resulted in most of the
change occurring due to Progressive Defoliation (figure 3) at approximately 67% of the change
area. The least amount of change occurred due to Avalanche, just under 2% of the change area
(table 3). In considering overall change within the NOCA study area, approximately 0.0043% of
the entire study area changed between 2007 and 2009. The combined percentage of change
occurring due to Progressive Defoliation equaled approximately 0.0029 % and Avalanche an
estimated 0.00006 % of the entire NOCA study area (table 4, figure 4). Using Google Earth and
TimeSync software for image interpretation will result different classifications in Random Forest
classification. The use of these programs is applicable when users have a prior knowledge of the
study area and experience in the initial assignment of agents of change.
Table 1. Table of 55 LandTrendr sample patches and assigned agent of change, with corresponding areas
within the NOCA study area.
Grid
Code
Change
Year
5767
2009
6308
2007
6415
6434
6484
Agent of
Change
Mass Movement
Area
(Hectares)
Grid
Code
Change
Year
1.26
11564
2007
0.90
12080
2007
2009
2009
2009
Progressive
Defoliation
Avalanche
Avalanche
Riparian
11.25
0.81
0.90
12154
12158
12159
2007
2008
2008
6588
2008
Mass Movement
2.07
12199
2007
6753
2007
Mass Movement
0.90
12231
2007
6822
2007
Avalanche
1.62
12311
2007
6845
2007
Mass Movement
1.62
12625
2007
6862
2008
Mass Movement
0.90
12701
2007
6863
2007
Riparian
1.08
12862
2007
6943
2007
Riparian
0.99
12872
2008
7004
2007
Riparian
0.90
12900
2007
8262
2007
Riparian
0.81
12940
2007
8936
2009
Tree Toppling
1.35
13356
2009
8973
9050
9850
9935
10275
10276
2009
2009
2008
2008
2008
2008
20.97
4.86
3.33
11.61
3.24
1.08
13374
13451
13533
13603
13610
13707
2007
2007
2007
2007
2007
2007
6562
2007
1.89
13723
2007
8006
2007
1.62
13758
2007
10490
2008
Mass Movement
Fire
Tree Toppling
Mass Movement
Riparian
Tree Toppling
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
0.99
13761
2007
10811
2008
Tree Toppling
1.35
13772
2007
10958
2008
Progressive
Defoliation
2.25
12229
2007
11008
2008
Avalanche
1.17
12757
11088
2009
Progressive
Defoliation
2.52
Agent of
Change
Progressive
Defoliation
Progressive
Defoliation
Fire
Mass Movement
Fire
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Tree Toppling
Progressive
Defoliation
Fire
Fire
Fire
Fire
Fire
Mass Movement
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Progressive
Defoliation
Fire
Progressive
Defoliation
Total Change for 55 sample patches:
2009
Area
(Hectares)
1.98
2.25
0.81
3.42
44.37
1.26
7.74
1.08
1.71
1.17
1.35
1.35
6.57
2.79
0.81
8.55
1.26
4.14
2.34
2.70
0.90
1.80
3.96
1.80
1.08
1.35
0.90
193.68
Hectares
LandTrendr Polygon
Study Area Boundary
Change patches
Fire Patches
Table 2. Table depicting the annual number of patches assigned to agents of change within the
NOCA study area.
Sample Data Assigned
Agent of Change
Avalanche
Fire
Mass Movement
Progressive Defoliation
Riparian
Tree Toppling
Total/Agent
Percent/Agent
39.82
Total/Year Percent/Year
14.85
7.67
70.38
36.34
43.65
22.54
46.98
24.26
7.92
4.09
9.9
5.11
Sample Change
Area:193.68
23.56
Avalanche
45
Area (Hectares)
40
Fire
35
30
Mass Movement
25
20
Progressive
Defoliation
Riparian
15
10
5
Tree Toppling
0
2007
2008
Year of Detected Change
2009
Figure 2. Annual area of change detected per change agent within the NOCA study area.
Table 3. Annual and overall number of patches assigned to each agent of change within the
NOCA study area.
Agent
Count
Avalanche 10
2007 1
2008 1
2009 8
Fire
2007
2008
2009
58
34
10
14
Agent
Count
Mass
Movement 56
2007 16
2008 28
2009 12
Agent
Count
Riparian
19
2007 11
2008 3
2009 5
Progressive
Defoliation
2007
2008
2009
Tree
Toppling
2007
2008
2009
249
138
29
82
20
3
13
4
140
120
100
80
60
Total
40
20
0
2007 2008 2009 2007 2008 2009 2007 2008 2009 2007 2008 2009 2007 2008 2009 2007 2008 2009
Avalanche
Fire
Mass Movement
Progressive
Defoliation
Riparian
Tree Toppling
Agent of Change
Figure 3. Annual amount of hectares assigned to agent of change within the NOCA study area.
Table 4. Overall and annual changes within the NOCA study area between 2007 and 2009.
Avalanche
0.9
1.08
15.21
Fire
140.2
22.5
21.51
Mass
Movement
24.03
55.62
27.54
Progressive
Defoliation
400.5
68.13
328.5
Riparian
14.85
2.88
24.12
Tree
Toppling
5.58
24.75
7.92
17.19
184.2
107.19
797.13
41.85
38.25
1.45
15.54
9.04
67.22
3.53
3.23
Percent
(%)/Year
49.42
14.75
35.82
Avalanche
Fire
350
Area (Hectares)
Total/
Year
586.08
174.96
424.8
300
Mass Movement
250
Progressive
Defoliation
Riparian
200
150
100
50
Tree Toppling
0
2007
2008
2009
Figure 4. Annual agents of change within the NOCA study area between 2007 and 2009
assigned using visual analysis.
Table 5. Annual percentage of change per change agent for the NOCA study area.
Year
Avalanche
2007
2008
2009
Total/Agent
3.2671E-06
3.9205E-06
5.5214E-05
6.2402E-05
5.0894E-04
8.1678E-05
7.8084E-05
6.6870E-04
8.7232E-05
2.0191E-04
9.9974E-05
3.8911E-04
1.4539E-03
2.4732E-04
1.1925E-03
2.8937E-03
Riparian
Tree
Toppling
Total/Year
5.3907E-05
1.0455E-05
8.7558E-05
1.5192E-04
2.0256E-05
8.9845E-05
2.8751E-05
1.3885E-04
2.1275E-03
6.3513E-04
1.5421E-03
4.3047E-03
Avalanche
Fire
Percent of NOCA
0.2
Mass
Movement
Progressive
Defoliation
Riparian
0.15
0.1
0.05
Tree Toppling
0
2007
2008
2009
Figure 5. Annual percentage of change per change agent for the NOCA study area.
DISCUSSION
Random forest variables used to classify the landscape patches into specific categories
were calculated prior to analysis. These variables include elements of geometry, landscape,
spectral, and time of change for each patch. The resulting Random Forest classification from the
initial assignment of agents of change to sample patches resulted in an out of bag error of
41.82% (table 1). This error signifies that a mere 58.18% of classes were accurately assigned to
appropriate agents of change based upon similar attribute data. Errors in assignments may have
resulted from noticing differences in imagery between 2007 and 2009, and applying assumption
based upon landscape position and closeness to fire region. For instance, several patches were reassigned because they did not appear to be in a region of fire occurrence, but after zooming out
from the change polygon fire was noted to have affected the general change area. Conversely,
many patches may have been incorrectly assigned as fire, when change may have affected the
area prior or precedent of fire occurrence. Notably, 1/3 of Mass Movement patches were
incorrectly assigned to Fire agents of change and 1/3 of Fire patches were assigned as other
agents (table 1). Generally, the assignment of agents of change should heavily rely on the change
in spectral values for a change polygon, which requires experience. Furthermore, the assignment
of agents should include consideration of the attribute data. For instance, a correlation between
slope position and occurrence of mass movements.
Nuisances in the classification results may have also resulted from a confusion between
different spectral value differences occurring throughout the change patches. For instance, the
assignment of agents of change for riparian areas resulted in an out-of-bag error of 66%, where
all incorrectly assigned patches would be more-accurately assigned to mass movement.
According the NCCN Landscape Dynamics Monitoring Protocol (2012), mass movements are
likely to occur due to debris flows from stream discharge, and should be classified in change
areas with a slope greater than 15%. In general riparian areas within the NOCA study area occur
at the base of large slopes or in relatively flat valleys (slope less than 15%). Even in considering
the landscape position and the wetness index of vegetation, riparian areas will have a greater
wetness index than surrounding areas. Therefore, spectral values for wetness should be compared
to surrounding landscape, and also take into account the landscape position relative values for
change.
The category of annual variability was not included in this analysis, however it is none
the less important in considering overall change within the NOCA study area. Annual variability
usually results from changes in snow or cloud cover, terrain shadows, or vegetation lifecycles
that are not removed from images during processing. This category of variability should be noted
because misclassification may occur due to the red, orange, or brown hues in the wetness index
of tasseled-cap imagery (Kennedy et al, 2012). Also, even though a particular region may be
classified as annual variability one year, the change patch may actually experience change in
another. In order to accurately assigned annual variably as an agent of change, the date and time
of acquisition must be considered (Kennedy et al, 2012) in order to determine such factors as sun
angle and climatic conditions. Considerations for annual variability must also be made in
examining change patch polygons in Google Earth.
Google Earth was a useful tool in examining change patch polygon landscape position
and closeness of region of fire occurrence. The use of Google Earth, however, should only be
considered as a tool for understanding general location of aspects within a study. For instance,
several change patch polygons were located in regions that experienced shadowing on the
Google Earth image. Therefore, most of the image interpretation should rely on TimeSync
software when using this method for analysis. TimeSync software is a tool that can help users
examine the spectral values for regions of interests, which is the main focus of remote sensing
processes for analysis. The magnification of regions is a useful aid in detecting change for a
small region, and computation on a large scale can be made possible thought other programs.
Large scale imagery would also be useful in ENVI software, however, Timesync was applicable
in this particular analysis. Comparisons between the software classification results could also be
made for a further in depth analysis. Overall, any remotely sensed analysis would be less timely
than field collection of data, but distinguishing between change agents would be easier in the
field.
LITERATURE CITED
Antonova, Natasha. 2014. Lab 3: Environmental Warfare in the Persian Gulf.
<http://staff.wwu.edu/antonon/envr442/ENVI/442_lab7_ENVI.html>
Antonova N et al. 2013. Landsat-based monitoring of landscape dynamics in the North Cascades
National Park Service Complex: 1985-2009. Natural Resource Data Series.
NPS/NCCN/NRDS2013/532. National Park Service. Fort Collins, Colorado.
Published Report-2198729
Kennedy, R.E, W.B. Cohen, A.A. Kirschbaum and E. Haunreiter. 2007. Protocol for Landsatbased monitoring of landscape dynamics and North Coast and Cascades Network Parks:
U.S. Geological Survey Techniques and methods 2-G1.
Kennedy, R.E., W.B. Cohen and T.A. Schroeder. 2007. Trajectory-based change detection for
automated characterization of forest disturbance dynamics. Remote Sensing of
Environment 110:370-386.