Geoscience Reference
In-Depth Information
freedom and generalized cross validation based on use of default criteria. As such,
the GAMs may have effectively already accounted for much of the variance
explained and inclusion of additional variables added little in terms of additional
model significance. Moreover, differences between the GAMs and MLR models,
although not large, typically added some 5-15 % in terms of variance explained,
which was a statistically significant improvement over the MLR models (0.01
>
p
<
0.05). In one case with landscape weighted variables, 76 % of variation in EPT
abundance was explained by a GAM using just a single predictor variable - the
amount of impervious cover within watersheds (Tables
4.1
and
4.2
).
Watershed size was not often selected as a significant predictor of the biotic
metrics in any of the model formulations, despite the area encompassed by the
watersheds ranging substantially (from 2 to 146,000 ha). Related, although there
were not many more cases in the catchment than in the aggregated watershed
scenarios, predictions at the catchment scale were consistently better than those
based on aggregating the watersheds. This finding suggests that the greater number
of smaller watersheds in the case of the catchment scenarios reflected greater
sensitivity of the biotic metrics to the land cover variables, even though the sample
sizes were only slightly greater (32 versus 27 cases).
Models of EPT abundance and richness performed consistently better than those
of the HBI metric, which is consistent with EPT representing sensitive taxa that
are better indicators of watershed impacts associated with urbanization than HBI
or other integrated indices of biotic integrity. Perhaps surprisingly, landscape
weighting did not consistently improve model predictions, and in some cases
actually reduced the variance explained by the different models. This finding is
consistent with the results of some of our previous analyses in the mid-Atlantic
region and may be due, in part, to the fact that impervious cover, which is
consistently selected as the most important predictor variable, is often directly
connected to the stream channel via storm drainage networks. In such cases,
deep-rooted forest cover, or other vegetation cover (grasslands or shrublands),
may be bypassed and landscape influences, in terms of buffering runoff volume
and pollutant contents, effectively minimized or obviated. Nonetheless, the MLR
models of EPT for the catchment scenarios selected tree cover as the primarily
predictor variable, despite impervious cover being consistently selected as the
primary predictor in nearly every other model formulation.
Only after landscape weighting did the amount of impervious cover get selected
as a significant predictor in these cases, which indicates that the weighting did in
fact effectively capture the buffering capacity of the landscape. Less variance was
explained in the models with landscape weighting, but this is because tree cover
was used in the weighting scheme and thus the relative importance of tree cover as
a predictor variable dropped out since its effects were already accounted for in the
landscape weighting scheme. Related, the HBI metrics showed some sensitivity to
landscape weighting in terms of changes in the selection of predictor variables,
although impervious cover remained the most important predictor. These findings
are consistent with EPT taxa richness and abundance being sensitive indicators of
urbanization impacts associated with impervious cover.
