PE&RS June 2016 Full - page 444

with differing relief and landforms. This result suggests that
physiography should be considered when using topographic
variables to map a landscape feature, such as palustrine wet-
lands. It should be noted that a direct comparison of variable
importance to prior studies that used a combination of terrain,
spectral, and other variables in different landscapes may be
inappropriate. Instead, this study is designed to simply inves-
tigate the conditional importance of the terrain variables inves-
tigated in this specific landscape for mapping the topographic
probability of palustrine wetland occurrence. Thus, direct
comparisons to other studies should be made with caution.
Predicting Wetlands Not Mapped in the NWI
The statistics for the single statewide model, combining PEM
and PFO/PSS wetlands, are shown in Table 4. This model, when
validated using a random sample from within the NWI, pro-
duced an AUC value of 0.986, in comparison to an AUC value
of 0.956 when validated using non-NWI wetlands. These AUC
values were shown to be statistically different at the 95 percent
confidence interval (p = <0.001). Thus, in summary, the model
trained with NWI wetlands was more accurate in predicting wet-
lands within the NWI than predicting wetlands outside of the
NWI. However, the AUC value for predicting outside of the NWI
was still well above 0.9, suggesting strong model performance.
The mean probability of occurrence for NWI wetland vali-
dation samples was 0.916 with a standard deviation of 0.171
(Table 4). For the non-NWI wetland validation samples the mean
probability of occurrence was 0.789 with a standard deviation of
0.261. The accuracy and AUC data all indicate that the model is
more accurate at predicting wetlands within the NWI compared
to non-NWI wetlands. Nevertheless, the summary statistics indi-
cate strong model performance for predicting non-NWI wetlands.
Figure 3 shows kernel density plots (i.e., probability
density functions) for wetland occurrence using the presence
and pseudo absence random samples from both the NWI and
non-NWI validation data. The figure shows clearly why the
AUC values are so high: most wetlands have modeled wetland
presence probabilities greater than approximately 70 percent,
while most non-wetlands have very low values, 20 percent or
less, with relatively few samples with probability values in
between. The lower maximum kernel density for the non-NWI
wetlands is notable, suggesting this data set comprises wet-
lands with characteristics that differ from those of the NWI.
The model trained using non-NWI data and tested against
non-NWI data resulted in a AUC value of 0.957, while the same
model tested with NWI data resulted in an AUC of 0.948, a
statistically different value (p = 0.0295). The fact that both the
NWI-trained and non-NWI-trained models each gave higher AUC
values when tested against other wetlands from the database
from which the training samples were selected, lends strong
support to the interpretation that the NWI and non-NWI wetland
databases, and their associated random forest models, do differ.
The reason for this decreased model performance when
predicting the probability of non-NWI wetlands is possi-
bly that the NWI includes almost all of the larger and more
distinctive wetlands (Tiner, 1997). The remaining unmapped
wetlands are likely small or narrow wetlands, and likely are
Figure 3. Kernel density plot showing distribution of probabilities of predicted palustrine wetland occurrence based on a model trained
with NWI data, for wetland presence and pseudo absence validated using NWI and non-NWI data.
T able 4. Comparison of Classification of Validation Data within NWI Mapped
Extents and Outside NWI Mapped Extents
Validation data source
Measure NWI Non-NWI
AUC 0.986 0.956
Mean of Wetland Probability 0.916 0.789
Standard Deviation of Wetland Probability 0.171 0.261
444 June 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING