Spectral information were taken from satellite imagery of the study... image was obtained from EROS Data Center (Sioux Falls, South... APPENDIX A

advertisement
Estimating insect distributions in Alaskan landscapes not covered in aerial surveys.
APPENDIX A
Modeling and Error Estimation Descriptions
GIS and Landsat Data
Spectral information were taken from satellite imagery of the study site. A Landsat-5 TM
image was obtained from EROS Data Center (Sioux Falls, South Dakota), while a MODIS
image was obtained from the Eric Johnson (Geospatial Coordinator, GIS/ALP, USDA
Forest Service, Juneau, Alaska). Both the Landsat-5 TM and MODIS imagery were
obtained for the study site during the month of August, 2007 coinciding with the field data
collection. The Landsat-5 TM imagery had seven spectral bands with a 30 m resolution,
while the MODIS imagery had six spectral bands with a resolution of 250 m (bands 1 and
2) and 500 m (bands 3-6). These latter four MODIS bands were resampled to a 250 m
spatial resolution using nearest neighbor techniques (Muukkonena and Heiskanenb, 2005).
Nearest neighbor resampling was selected due to quicker computer processing time as
compared to other interpolation methods. In addition, nearest neighbor interpolation better
maintains original reflectance values while providing sufficient accuracy and reduced
potential introduction of unwanted geometric distortions in areas with no ground control
points to provide precise control (Muukkonena and Heiskanenb, 2005).
Topographic data were taken from a digital elevation model (DEM). The DEM was
obtained from the National Elevation Dataset (NED) as a seamless ArcInfo (ESRI, 1995)
grid at a 90 m resolution (U.S. Geological Survey (USGS), Gesch et al., 2002; Rabus et al.,
2003). The DEM was resampled to both a 30 m and 250 m spatial resolution using bilinear
techniques (Edenius et al., 2003), producing a more continuous surface reflecting gradual
changes in elevation at both a 30 m and 250 m spatial resolution. GIS grids of elevation,
slope, and aspect were derived from the DEM’s using ArcView (ESRI, 1998). Values for
all grid layers of information were derived for each sample point station using Avenue
(ESRI, 1998) code.
Modeling Vegetation Types and Distribution of Aspen Leaf Miner
Binary classification trees were used to describe the presence/absence of a given
vegetation type infested and non-infested as a function of topographic data and the
satellite imagery (Joy et al., 2003). A 10-fold cross-validation procedure was used to
identify the best tree size to minimize the prediction classification error. The pruned
classification trees were then used to generate GIS layers showing the spatial distribution
of the various vegetation types at a 30 m resolution for the Landsat-5 TM imagery and at
a 250 m resolution for the MODIS imagery. This was accomplished by passing the
appropriate GIS layers of topographic data and satellite imagery through the pruned
classification trees. Receiver Operating Curves (ROC) were used to assess the predictive
accuracy of the binary models (Boyce et al., 2002).
1
An ROC curve is a plot of the true positive rate against the false positive rate for different
possible outcomes of a model. An ROC curve demonstrates several things: 1) the tradeoff between sensitivity and specificity; 2) the closer the curve is to the left-hand border
and the top border of the ROC space, the more accurate the model; 3) the closure the
curve comes to the 45-degree diagonal of the ROC space, the less accurate the model; 4)
the slope of the tangent line at a give point in the ROC curve gives the likelihood ratio for
that value of the model; and 5) the area under the curve is a measure of test accuracy
(Boyce et al., 2002). This last characteristic, the area under the curve measures
discrimination, or the ability of the model to correctly classify a sample point as to its
vegetation type.
Modeling Forest Characteristics
Modeling of basal area and canopy closure was accomplished in two stages using
procedures developed by Reich et al. (2004a). In the first stage, multiple regression analysis
was used to describe the coarse-scale variability in the stand structure as a function of
elevation, slope, aspect and the Landsat-5 TM or MODIS bands. For each component of
forest structure, a stepwise procedure was used to identify the best subset of independent
variables to include in regression models. Semi-variograms were used to evaluate spatial
dependencies among residuals from the various models. If the residuals exhibited spatial
dependencies, a generalized least squares model was used to estimate the regression
coefficients associated with the regression models (Reich and Davis, 1998). In the second
stage, a tree-based stratification design was used to enhance the estimation process of the
small-scale spatial variability. With this approach, sample units (i.e., pixel of a satellite
image) are classified with respect to predictions of error attributes into homogeneous
classes, and the classes are then used as strata in the stratified analysis. Independent
variables considered in the binary regression trees (Breiman et al., 1984) included elevation,
slope, aspect and Landsat-5 TM or MODIS bands. A decision rule was used to identify a
tree size that minimized the error in estimating the variance of the mean response and
prediction uncertainties at new spatial locations.
The effectiveness of the final models were evaluated using a goodness-of-prediction
statistic (G) (Agterberg, 1984; Kravchenka and Bullock, 1999; Guisan and Zimmermann,
2000; Schloeder et al., 2001). The G-value measures how effective a prediction might be
relative to that which could have been derived using the sample mean (Agterberg, 1984).
A G-value equal to 1 indicates perfect prediction, a positive value indicates a more
reliable model than if one had used the sample mean, and a negative value indicates a less
reliable model than if one had used the sample mean.
Grids of basal area and canopy closure were generated for the best fitting regression
models. Similarly, grids representing the error associated with each regression model
were generated by passing each grid for the appropriate independent variables through the
regression trees. The final predicted surfaces were obtained from the sum of the two
grids.
2
Model Evaluation
A 10-fold cross-validation was used to estimate the prediction error for each model
(Efron and Tibshrani, 1993; Stone, 1974). Data were split into K=10 parts consisting of
approximately 21 sample points and models were fitted to the remaining K-1=9 parts of
the data. The fitted model was used to predict the part of the data removed from the
modeling process. This process was repeated 10 times so that each weather station was
excluded from the model fitting step and its response predicted. Prediction errors were
obtained from the predicted minus actual values.
Various measures of prediction error were calculated to evaluate the effectiveness of the
models. Prediction bias (Williams, 1997) was calculated for each validation data set as a
percentage of the true value and accuracy (Kravchenko and Bullock, 1999; Schloeder et
al., 2001) was measured by the mean absolute error (MAE), which is a measure of the
sum of absolute residuals (i.e., actual minus predicted) and the root mean squared error
(RMSE). Small MAE values indicate a model with few errors, while small values of
RMSE indicate more accurate predictions on a point-by-point basis (Schloeder et al.,
2001).
Estimation uncertainty in the models (Reich et al., 2004a) was calculated as the
estimation error variance (EEV), σˆ i2* for each observation in the data set. The
consistency between the EEV and the observed estimation errors (i.e., true errors) was
calculated using the standard mean squared error (SMSE) (Hevesi et al., 1992). EEVs
were assumed consistent with true errors if the SMSE fell within the interval
[1 ± 2(2 n ) ]
1/ 2
(Hevesi et al., 1992). The EEVs were also used to construct 95%
confidence intervals around individual estimates. Coverage rates were calculated as the
proportion of individual confidence intervals that contained the true value.
Models were developed to predict the spatial distribution of Aspen leaf miner in the study
area using satellite imagery. The results are summarized in the Table below.
Results -- Modeling Vegetation Types and Distribution of Aspen Leaf Minor
All aspen stands were heavily infected and because of the difficulty in locating stands
with low to moderate levels of infections it was difficult to discriminate between birch
and non-infected aspen stands and so for the analysis, these stands were combined. The
Landsat-5 TM model was able to discriminate between all land cover types, except open
areas, with an overall accuracy of 81% (Table 1). A second measure of accuracy is
provided by the area under the curve of the ROC curve. An area of 1 represents perfect
prediction, while an area of 0.5 represents a worthless prediction. A rough guide for
classifying the accuracy of the model is the traditional academic point system
● 0.90-1 = excellent (A)
3
● 0.80-0.90 = good (B)
● 0.70-0.80 = fair (C)
● 0.60-0.70 = poor (D)
● 0.50-0.60 = fail (E)
The area under the curve for non-infected aspen and spruce was 0.89 and the model
would be considered almost excellent in the ability to discriminate between these two
vegetation types. Infested aspen had an area under the curve of 0.85 so the model would
be considered good at discriminating this vegetation type. The model was only fair (area
under the curve – 0.78) in its ability to discriminate open areas.
Vegetation Type
Accuracy
Area Under the
Curve1
0.85
0.89
ROC
Ranking1
Good
Good
Infected Aspen
0.80
Birch/Non-infected
0.89
Aspen
Spruce
0.82
0.89
Good
Open Areas
0.59
0.78
Fair
Overall
0.81
1
AUC = 1-0.90 – excellent; 0.80 – 0.89 – good; 0.70-0.79 – fair; 0.60-0.69 – poor;
0.50 – 0.59 – failure.
Table 1. Accuracy assessment of Landsat-5 TM imagery in identifying vegetation types
on the study site near Fairbanks, Alaska.
Figure 2 displays the predicted spatial distribution of the vegetation types and associated
ROC curve, while Table 3 summarizes the estimated areas associated with each of the
vegetation types. Spruce dominated the study site (45%) followed by birch/non-infected
aspen stands (31%) and then infected aspen stands (22%). Open areas covered only 2% of
the study site.
The MODIS model had a lower overall accuracy (69%) compared to the model based on
the Landst-5 TM imagery (Table 3). Spruce (78%) and birch/non-infected aspen stands
(73%) had the highest level of accuracy, while infected aspen stands (60%) and open
areas (59%) had the lowest accuracy. Fig.3 displays the predicted spatial distribution of
the vegetation types and associated ROC curve. The area under the curve indicted that the
model was only fair in its ability to discriminate among the various vegetation types,
except for spruce in which the model was able to do a good job in discriminating this
vegetation type.
Table 4 summarizes the estimated areas associated with each of the vegetation types. The
MODIS model over estimated the area in spruce and infected aspen stands and under
estimated the area covered by birch/non-infected aspen stands when compared to the
Landsat-5 TM model. The area covered in open areas was similar for both models.
4
1.0
Vegetation Types of Study
Area Near Fairbanks, Alaska
Birch
Aspen
0.6
Open
0.4
TPR (sensitivity)
0.8
Spruce
N
i sc
od
ti
i na
r im
(r
on
m
do
an
e ss
gu
)
0.2
Better
Legend
Worse
Roads
4
0.0
Water
Aspen
0
1,125 2,250
4,500
6,750
9,000
Miles
Birch/Non-infected aspen
Open areas
0.0
0.2
0.4
0.6
0.8
1.0
FPR (1-specificity)
Spruce
Fig. 2. Spatial distribution of vegetation types predicted using Landsat TM imagery
on a study site near Fairbanks, Alaska, and associated ROC curve.
Table 2. Estimated areas associated with the Landsat-5 TM model for the
various vegetation types on the study site near Fairbanks, Alaska.
Vegetation Type
Area (ha)
Percent
Infected Aspen
4,433
22.2
Birch/Non-infected Aspen
6,216
31.1
363
1.8
Spruce
8,979
44.9
Total
19,981
Open Areas
Table 3. Accuracy assessment of MODIS imagery in identifying vegetation types on
the study site near Fairbanks, Alaska.
Tree Cover
Accuracy
Area Under the
ROC
1
Curve
Ranking1
Infected Aspen
0.60
0.76
Fair
Birch/Non-infected
0.73
0.77
Fair
Aspen
Spruce
0.78
0.82
Good
Open Areas
0.59
0.78
Fair
Overall
0.69
5
1
AUC = 1-0.90 – excellent; 0.80 – 0.89 – good; 0.70-0.79 – fair; 0.60-0.69 – poor;
0.50 – 0.59 – failure.
1.0
Vegetation Types of Study
Area Near Fairbanks, Alaska
0.8
Spruce
0.6
Aspen
0.4
TPR(sensitivity)
Birch
Open
d
No
i na
r im
i sc
ti o
ra
n(
om
nd
e ss
gu
)
Better
Roads
Water
Aspen
0
1,125 2,250
4,500
6,750
9,000
Miles
Worse
0.0
4
0.2
Legend
Birch/Non-infected aspen
Open areas
Spruce
0.0
0.2
0.4
0.6
0.8
1.0
FPR(1-specificity)
Fig. 3. Spatial distribution of vegetation types predicted using MODIS imagery on a
study site near Fairbanks, Alaska, and associated ROC curve.
Table 4. Estimated areas associated with the MODIS model for the various
vegetation types on the study site near Fairbanks, Alaska.
Vegetation Type
Infected Aspen
Area (ha)
Percent
6,306
30.6
3025
14.7
219
1.1
11,037
53.6
Birch/Non-infected Aspen
Open Areas
Spruce
Total
20,587
Results -- Modeling Forest Characteristics
Canopy closure and basal were linearly correlated with the topographic (elevation and
slope) and satellite imagery, and these linear relationships varied significantly among
vegetation types. All of the topographic, satellite imagery and vegetation type data? were
also used in regression trees to describe the error in one or more of the regression models.
The decision rule used to identify the optimal tree size identified 9 strata for the canopy
closure model using the Landsat-5 TM data and 9 for the basal model and with a minsize
6
of 60 observations. The basal area model accounted for 53% of the variability observed
on the sample plots while the canopy model had a G-statistic of 42% (Table 5). In
contrast, the models for canopy closure and basal area using the MODIS data both had
strata sizes of 8, respectively and with a minsize of 60 observations. The canopy model
had a G-statistic of 58% and 52% for the basal area model (Table 5). Sample plots with
the same canopy closure or basal area can have different spectral properties depending on,
for example the vegetation type, stocking level and average tree size thus, making it more
difficult to model using the spectral reflectance and terrain data. This results in a coarser
partitioning of the data in a small number of large strata. This coarser partitioning of the
data results in a lower G-statistic for these two models. Fig. 4 and 5 compare the final
models for canopy closure and basal area using the two sets of satellite imagery
The confidence and prediction coverage rates for all models were close to the 95%
nominal coverage rate. The standardized mean squared errors were close to their
expected value of unity indicating the variance estimates were consistent with the true
errors. The standardized mean squared errors ranged from 0.91 to 1.16. Figure 6 display
the histograms of the prediction errors from the 10-fold cross-validation of the final
models. The histograms and plots of the predicted vs. observed values (not shown) did
not show any trends suggesting any systematic bias in the models.
Canopy Closure on the Study
Area Near Fairbanks, Alaska
Canopy Closure on the Study
Area Near Fairbanks, Alaska
Roads
Roads
Water
Water
Percent
Percent
0
0
0 - 20
0 - 20
20-40
4
20-40
4
40-60
0
0.5
1
2
3
60-80
4
Miles
80-100
40-60
0
0.5
1
2
3
60-80
4
Miles
Fig. 4. Comparison of the predicted spatial distribution of canopy closure on
the study site near Fairbanks, Alaska using Landstat-5 TM (left) and
MODIS data (right).
7
80-100
Table 5. Overall model performance (G-statistic) of the multiple
regression models (LS) and the use of regression trees (RT) in
describing the errors in the LS models of forest structure using
Landsat-5 TM and MODIS imagery on a study site near Fairbanks,
Alaska.
Model 1
1
LS
LS + RT
Canopy Closure-L (%)
0.182
0.453
Basal Area-L (m2 ha-1)
0.422
0.526
Canopy Closure-M (%)
0.448.
0.577
Basal Area-M (m2 ha-1)
0.395
0.519
L=Landsat-5 TM; M=MODIS imagery
Prediction bias was nominal (Table 6) for all models. Minimum, maximum, and quartile
values showed that estimated and observed value distributions were similar for all
models. Estimation errors for the canopy closure and basal area had a similar spread in
the estimation errors for both set of satellite imagery. The mean estimation errors did not
differ significantly from zero (p-value > 0.050). The MAE was smaller than the RMSE
for all models indicating that in general, the models are more accurate in predicting
regional or global means than on a point-by-point basis.
Basal Area on the Study
Area Near Fairbanks, Alaska
Basal Area on the Study
Area Near Fairbanks, Alaska
Roads
Roads
Water
2
m ha
Water
-1
2
m ha
0
0 - 40
0 - 40
40-50
40-50
50-100
50-100
100-150
4
100-150
4
150-200
0
0.5
1
2
3
200-300
4
Miles
-1
0
300-500
150-200
0
0.5
1
2
3
200-300
4
Miles
Fig. 5. Comparison of the predicted spatial distribution of basal area on the
study site near Fairbanks, Alaska using Landstat-5 TM (left) and
MODIS data (right).
8
300-500
Table 6. Summary statistics of estimation errors of forest structure using
Landsat-5 TM and MODIS imagery for a study area near Fairbanks,
Alaska based on the 10-fold cross-validation.
N
188
188
205
Basal
Area –M2
(m2 ha1)
205
Mean
1.34
0.96
-2.35
0.63
IQR
24.26
86.09
27.04
109.09
MAE
16.96
55.10
18.83
65.42
RMSE
22.60
71.88
25.69
81.83
SMSEM
0.92
0.95
0.91
0.91
SMSEP
1.16
0.92
0.99
1.06
CRM
0.97
0.95
0.96
0.96
CRP
0.93
0.96
0.92
0.95
Statistic
1
Canopy
Basal
2
Closure- L Area –L2
(%)
(m2 ha-1)
1
Canopy
Closure-M2
(%)
IQR = interquartile range, MAE = mean absolute error, RMSE = root mean
square error, SMSE = standardized mean square error for the mean response
(M) and for prediction; CR = coverage rate for the mean response (M) and
prediction (P).
2
L=Landsat-5 TM; M=MODIS image
9
Table 7. Summary statistics of observed and estimated forest structure using Landsat-5 TM and MODIS imagery for a study site
near Fairbanks, Alaska from 10-fold cross-validation.
Canopy Closure-L1 (%)
Basal Area-L1 (m2 ha-1)
Statistic
Obs.
Est.
N
188
188
Mean
67.24
65.9
Std. Dev.
22.98
14.22
CV%
34.17
21.58
Minimum
0.0
16.01
First quantile
59.48
64.24
Median
72.18
68.53
Third quantile
84.38
73.2
Maximum
96.35
89.07
Bias%
1.99
1
L=Landsat-5 TM; M=MODIS imagery
Obs.
188
133
80.95
60.87
0.0
70.0
120.0
180.0
370.0
Est.
188
132
63.82
48.34
0.0
89.07
136.70
180.0
272.2
0.75
10
Canopy Closure-M1
(%)
Obs.
Est.
205
205
61.77
64.12
22.98
19.84
34.17
30.94
0.0
0.0
52.0
61.97
71.0
70.65
83.0
75.87
96.0
97.05
-3.80
Basal Area-M1 (m2 ha1
)
Obs.
Est.
205
205
133
120.7
80.95
64.07
60.87
53.09
0.0
0.0
70.0
73.5
120.0
124.2
180.0
176.0
370.0
266.3
9.25
60
b
40
30
20
Number Sample Plots
40
0
0
10
20
Number Sample Plots
60
50
a
-50
0
50
-200
-100
Residuals
100
200
100
200
Residuals
d
30
20
Number Sample Plots
40
0
0
10
20
Number Sample Plots
60
40
50
80
c
0
-100
-50
0
50
-200
Residuals
-100
0
Residuals
Fig. 6. Histograms of prediction errors from a 10-fold cross-validation of the canopy model
(a and c) and basal area model (b and d) using Landsat-5 TM (a nd b) and MODIS (c and d)
satellite imagery.
11
Download