Geoscience Reference
In-Depth Information
Fig. 4.4 A graph showing the output predicted vs. observed number of EPT for the aggregated
training watersheds in southern New England. These results are from an MLR model that did not
utilize landscape weighting of the land cover variables
Table 4.3 Tabular results
comparing general statistics
for predicted vs. modeled
watersheds for nEPT scenario
Land cover variables
% ISA % Tree % Crop % Grass
Predicted
Min
0.01
10.42
0.00
0.10
Max
53.75
91.15
31.03
52.21
Mean
6.16
64.44
1.66
7.86
Std dev
8.82
16.69
3.23
6.84
Modeled
Min
0.24
8.04
0.00
1.59
Max
41.75
87.66
1.60
10.75
Mean
12.25
55.50
0.51
4.98
Std dev
11.53
19.01
0.43
2.88
each state individually. Splitting the HBI data this way limited the number of
samples needed for robust GAM or MLR models of the landscape weighted
watersheds in Massachusetts. Furthermore, low sample sizes for each of the
tested scenarios precluded the use of other data intensive approaches, such as
decision-tree models, which have performed well in other studies of this nature
(e.g. Goetz and Fiske
2008
)(Fig.
4.4
).
Although the GAM results produced higher R
2
values overall, they also were
less consistent. The models were run several times for each scenario, to iterate on
variable selection and generate a different data subset for testing and validating the
models. The MLR results, on the other hand, were consistently more robust.
Considering these findings, we used the MLR models, and watersheds of southern
New England (with land cover relative to those we used to build the models -
Table
4.3
), to produce a map of predicted nEPT (e.g. Fig.
4.5
). This same model
