Note: TWI: topographic wetness index, Bare is bare ground cover; NDMI:
normalized differenced moisture index, NDVI: normalized differenced
vegetation index, EVI: enhanced vegetation index, dNBR: differenced
normalized burn ratio, and MTBS: monitoring trends in burn severity
database.
Topographic metrics included watershed area, mean slope, mean elevation,
riparian extent, and mean topographic wetness index (TWI) (Table 3).
Slope, elevation, and TWI were derived from the 10-m DEM using Whitebox
tools (Lindsay, 2020; Wu, 2021) and summarized as watershed means. We
used a physical definition of the riparian corridor that included pixels
<2 m above the stream surface elevation (sensu Jencso
et al., 2010) and calculated riparian extent as the total riparian
corridor area divided by UAA of each sampling point. This approach
differs from an earlier estimate of the extent of riparian vegetation in
these watersheds (Rhoades et al., 2019).
We characterized vegetation condition using normalized differenced
vegetation index (NDVI), normalized differenced moisture index (NDMI),
and enhanced vegetation index (EVI). We obtained mean June 2018
vegetation indices from Landsat using Climate Engine (Huntington et al.,
2017) to match the vegetation characterization with the timing of our
stream sampling. We also included 2018 fractional land cover estimates
derived from satellite imagery that was extensively calibrated across
the Western US and estimated the proportion of each Landsat pixel
covered by trees, shrubs, and bare ground (Allred et al., 2021).
Mean differenced normalized burn ratio (dNBR, a measure of burn
severity) and burn extent were calculated for the area contributing to
each sampling point. These fire metrics represent immediate post-fire
impacts by differencing one pre-fire (8/24/2001) and one post-fire
(8/14/2003) Landsat image. dNBR was then classified into categorical
burn severity as follows: -150-140 unburned; 140-211 low severity;
211-350 moderate severity, 350-953 high severity (Eidenshink et al.,
2009). Low severity fire tends to leave tree canopies largely unaltered
whereas high severity fire typically causes complete consumption of
surface organic matter and canopy foliage (Parsons et al., 2010).
Wildfire severity varies spatially across topographic, vegetation (i.e.,
fuel composition, arrangement, condition), and weather gradients (Taylor
et al., 2021) which creates mosaics of post-fire vegetation structure
and composition that vary at scales finer than mapped severity patches
(Lentile et al., 2007). To characterize the spatial burn patterning of
each watershed, we calculated burn extent, patch size, patch radius, and
patch density by severity (Table 2). Burn extent reflects the proportion
of watershed area that was burned by each severity class. All patch
metrics were calculated with the landscape metrics package (Hesselbarth
et al., 2019) in R Studio which defines contiguous cells belonging to
the same burn severity class. For each watershed, we determined patch
area and calculated patch radius as the mean distance from each cell in
a patch to its centroid, and patch density as the number of patches
divided by watershed UAA.
2.4 Statistical modeling
We used statistical models to evaluate the degree to which topographic,
vegetation, and fire variables and flow-connected distance control
post-fire stream water chemistry – specifically, concentrations of
Na+ and NO3-.
Concentration data were log-transformed to improve normality and a
correlation analysis removed redundant predictor variables with a
correlation >0.90 (Figure 2, Table 2). To identify the
top-performing Na+ and
NO3- models, we went through a
two-step model selection process (sensu McManus et al., 2020;
Rodríguez-González et al., 2019). First, we identified which landscape
characteristics best predicted stream Na+ and
NO3- using linear mixed model
selection (Supplemental Table 1). The Na+ and
NO3- models with the lowest maximum
likelihood estimate of Akaike’s Information Criteria (AIC) then
progressed to the second phase of model selection where we compared
spatial autocorrelation approaches. We initially ran multiple linear
regression (MLR) models which use landscape characteristics to predict
observed water chemistry at each sampling location. The predictor
variables are spatially explicit given that they characterize the area
contributing to a specific sampling point, but MLR models assume
independence between stream water samples. We then compared MLR models
to spatial stream network (SSN) models that jointly consider landscape
and stream network characteristics. This approach captures spatial
effects beyond those directly attributable to predictor variables by
accounting for flow-connection (Isaak et al., 2014). MLR and SSN model
performance was compared through iterative leave-one-out
cross-validation. Observations at sampling points were removed one at a
time and the model was used to predict each of the removed values along
with the prediction standard error (Ver Hoef & Peterson, 2020). The
model with the lowest AIC and root mean square prediction error (RMSPE)
was selected for subsequent analyses.