Note: TWI is topographic wetness index, NDMI is normalized difference
moisture index, AIC is Akaike’s information criteria, and RMSPE is root
mean square prediction error.
Fire and vegetation variables generally had stronger correlations
(0.15-0.67) with log[NO3-] than
topographic variables (0.03 – 0.32) (Table 3). Linear mixed model
selection identified watershed area, riparian extent, TWI, and NDMI as
the best predictors of log[NO3-]
(Supplemental Table 1). Stream NO3-was positively related to riparian extent and TWI, but negatively
related to watershed area and NDMI (Figure 2). Mean NDMI had the
strongest correlation with
log[NO3-] (Figure 2). In the
NO3- MLR model, the selected predictor
variables, with the exception of riparian extent, were significant and
accounted for 51.4% of the variance in
log[NO3-] (Table 4). In the
NO3- SSN model, TWI and NDMI were the
only significant predictor variables and the predictors explained 36%
of variation in log[NO3-] (Table
4).
Topographic variables had weak correlations (<0.32) with both
stream Na+ and NO3-(Table 3). Vegetation predictors generally had much stronger
correlations with NO3- compared to
Na+, with the exception of shrub cover (Table 3). Burn
variables had slightly higher correlations with
NO3- compared to Na+(Table 3). All predictor variables that were selected through linear
mixed model selection were weakly correlated with water chemistry
(<0.33) (Supplemental Figure 2). The one exception was a
strong inverse relationship between mean NDMI and stream
NO3- which had a correlation
coefficient of -0.67 (Supplemental Figure 2).
3.3 Stream network controls on
Na+ and NO3-
In the Na+ SSN model, a majority of variation (53.1%)
in log[Na+] was explained by flow-connected
autocorrelation (Table 4). Na+ exhibited strong
positive autocorrelation where semivariance was low at short lag
distances, but increased with distance (Figure 3). When flow-connected
autocovariance was modeled with a spherical fit, Na+had a nugget of 0.001, sill of 0.029, and range of 3700 m (Figure 3).
The low nugget suggests that our sampling adequately captured
variability at small spatial scales and that there is relatively little
unexplained variation. The low sill reflects the low overall variance in
streamwater Na+ concentrations. The range indicates
that samples that are > 3700 m apart are no longer
correlated.
In the NO3- SSN model, flow-connected
autocorrelation explained 41.5% of variation in
log[NO3-] (Table 4). Stream
NO3- had high semivariance across all
flow-connected distances, though semivariance peaked at intermediate lag
distances (1000-5000 m) (Figure 3). When flow-connected autocovariance
was modeled with an exponential fit,
NO3- had a nugget of 0.385, sill of
0.708, and range of 8800 m which is equal to our maximum sampling
distance (Figure 3). The large nugget and sill values are consistent
with the substantial unexplained variance and high overall variance in
stream NO3- concentrations. The lowest
semivariance in NO3- is still greater
than the maximum Na+ semivariance (Figure 3).