Geoscience Reference
In-Depth Information
distance weighting scheme is a distance to stream channel, where for each 30 m cell
within the watershed distance is calculated to the closest cell that defines the stream
channel. These values are scaled between 0 and 1, with 1 representing the greatest
possible distance. The second distance metric incorporates the amount of tree cover
within each pixel (30 m grid cell) of the watershed.
Originally containing values of 0-100%, the tree cover layer was also scaled
between 0 and 1. The inverse distance is then calculated for both of these
components, providing two data layers, one depicting the inverse distance to
stream, with values closest to the stream being the highest (at or near 1), and the
values furthest from the stream being the lowest (at or near 0). These two data
layers are combined (multiplied) to provide a scaled landscape weighted cost
surface that reflects the potential buffering capacity of tree cover on both over-
land and, to a lesser extent depending on rooting depth, subsurface flow.
Each of the other land cover classes (impervious, grass and crop cover) were
then multiplied by the landscape weighting scheme. Results using this distance-
weighting scheme for each watershed were compared to results without using
distance weighting. Analyses was conducted using ESRI
®
GIS software, making
use of Python and Arc Macro scripting languages to summarize the data for further
statistical analysis.
4.3.3 Statistical Analyses and Predictions
A stepwise Multiple Linear Regression (MLR) and General Additive Model
(GAM) was used to test the relationship between the biotic metrics (EPT and
HBI) and the land cover variables. In addition to the NLCD land cover layers
(percent impervious, tree cover, grassland, and cropland), predictor variables
included watershed size and the landscape weighted transformations of the
land cover variables. The response variables were the stream biotic metrics:
HBI, EPT abundance and EPT richness (Fig.
4.2
). The same predictor and
response variables were used for both the landscape weighted and the non-
weighted tests.
As in previous research, a forward stepwise MLR and GAM was used to
predict stream biota indictors from the land cover metrics. These procedures
allowed us to train a linear model on a portion of the data (90%) while withholding
a selection of the data (10%) for cross validation. Predictor variables were itera-
tively selected based on their relative power in explaining variance within the
response variable. Finally, the best fit models were used to create a map of
predicted stream biotic quality for each HUC12 watershed, which were selected
because they encompassed a comparable total watershed area (size) as those for
which HBI and EPT metrics were derived and aggregated.
