Discussion
Comparison with the Varhola et al. (2010)
Framework
Maximum snow accumulation has a negative relationship with increasing
fVEG using our “open reference site” lidar-based approach, which
aligns with the meta-analysis of Varhola et al. (2010). Our approach
takes advantage of the lidar point density to compare large scale (30-m)
snow distributions with a nearby open reference location (Figure A5),
which controls for factors such as precipitation and terrain. However, a
limitation of this approach is that the “open reference” area
implicitly depends on the amount and arrangement of vegetation (e.g. not
all open 1-m areas are the same). Using the flights that best
approximate maximum snow accumulation across three years (i.e. the least
ablation at lower elevations), we find a linear trend in relative snow
accumulation between fully open and 100% fVEG, with differences of 23%
in 2008 and 37% in 2016 (R2 of 0.72 and 0.62,
respectively). These slopes are either similar or smaller than those
from Varhola et al. (2010), who calculated a slope of 0.37. We see
stronger relationships in February 2008 but a lower slope (Figure 7),
potentially due to the lack of ablation at lower elevations as compared
with March 2016. More lower elevation ablation may accentuate DSnowAc
differences because under canopy areas are expected to ablate earlier
than open areas in this warmer environment . The correspondence between
2008 and 2016 patterns in DSnowAc with fVEG support the “open reference
site” approach and the resulting process inferences we describe below.
A random forest model demonstrates that fVEG accounts for the most
variation of any predictor variable (fVEG, openness, LAI’, northness,
elevation) in two of the accumulation models. The RF results show weaker
effects when fVEG is below ~30%. The shallower slopes
in the RF partial plots (Figure 9; ~10% change in
DSnowAc for 100% change in fVEG) compared with the linear relationships
suggest the importance of factors besides fVEG (discussed more in
Section 4.2).
Our “open reference site” approach predicts greater ablation with
increasing fVEG, and thus reductions in ablation rates with canopy
removal, which is the opposite of the relationships found by Varhola et
al. (2010). However, our results are consistent with process-based
studies in warmer climates, where high levels of longwave radiation from
trees may overcome the decreased solar radiation in dense forests,
especially on shaded north facing slopes Our results demonstrate a need
to refine the Varhola et al. (2010) model of ablation for different
climates. Steep north-facing slopes may not receive as much increase in
solar radiation after thinning compared to flat areas, which would
strengthen the signal of higher ablation with higher fVEG in these
areas. Greater under-canopy ablation occurs from a combination of the
timing of the ablation season (March-May, when solar irradiance is
relatively low) and trees that emit enough longwave radiation to
dominate energy fluxes . We find an average of ~15%
change in ablation for 100% change in fVEG for the March-April and
April-May periods (R2 of 0.13 and 0.34, respectively
in Figure 8). Partial plots in the RF model also show positive
relationships with fVEG and ablation rate but indicate that there is
little effect when fVEG is below ~30% (Figure A13).
Ablation rates calculated between two lidar flights are different than
season-long ablation metrics used by Varhola et al. (2010). In
particular, the lidar-based ablation metrics are affected by snow
accumulation during the ablation season, variation in precipitation
amount, cold content, energy inputs, and other factors that are highly
spatially variable (Figure A3). We attempt to mitigate these issues by
using early season (lower elevation) and late season (higher elevation)
ablation windows and controlling for initial SWE in each case. However,
opportunities exist to improve our method to better utilize ablation
information by including distributed snow disappearance estimates with
modeling or optical remote sensing. An increasing number of open-source,
multi-temporal lidar datasets, collected by ASO for example, in recent
years presents an opportunity to repeat these methods across the Sierra
Nevada and into domains, like Colorado and Utah, that may show different
results due to a variety of energy balance dynamics. Importantly, to be
the most useful for this method, these flights must be paired with
timely, high point-density, snow-off lidar as well.
Role of Finer-Scale Forest Structure and Terrain on Snow
Dynamics
The predictor importance shows that fVEG, LAI’, and openness are more
important than terrain metrics (elevation, northness, and eastness) for
predicting our DSnowAc metric in the RF model (Figure A9). The
importance of these vegetation metrics was clearly shown by our “open
reference site” approach since analyses of raw snow accumulation data
would fail to disentangle the role of vegetation from the strong
influence of elevation and snowfall spatial patterns (Figure 4 and
Figure 5). The partial plots from the RF model consistently show a
non-linear effect on accumulation from LAI’ and openness, as compared to
a more linear relation with fVEG. Overall, our results imply that
vegetation has the largest role in reducing snow accumulation when fVEG
is high (>0.6), openness is low (vegetation height ≥ canopy
gap diameter), and LAI’ is high (>0.25). Inversely, these
results indicate that a canopy gap diameter to vegetation height ratio
<1 optimizes accumulation, in general support of earlier
studies that demonstrated maximum shading across the gap when nearby
canopy height is comparable to the gap diameter (i.e. openness ratios
<1) . Conversely, partial plots for terrain effects on
accumulation and ablation have shallower slopes, with mixed slopes
across flights suggesting weaker controls (FigureA14-A18). Our results
match Krogh et al. (2020) and Lewis et al., (2023), who used the
SnowPALM model to conclude that lower-elevation, south-facing slopes
experience post-thinning increases in accumulation.
We also show the importance of fine-scale forest structure controlling
ablation rates, and promoting snow retention following thinning, despite
limitations in our random forest analysis. Because of differences in
energy environments, elevation is important for predicting DSnowAb,
whereas fVEG, openness, and LAI’ are more important for the early
ablation season metric. Both early and late season ablation estimates
indicate a combined impact of fVEG (lowest ablation at
<~50% canopy cover) and openness (lowest
ablation where gap diameter is greater than tree height) to promote
retention. The similar response of snow accumulation and ablation
suggests that forest thinning in dense areas – especially if the
thinning creates large gaps rather than uniform density reduction – is
likely to slow ablation while increasing accumulation, thereby
increasing overall snow retention.
Despite the importance of a coarse (simpler) fVEG metric based only on
tree spatial coverage, we find that metrics accounting for vertical
lidar point density (LAI’) and horizontal and vertical canopy
information (openness) help refine predictions of snow-forest
interactions. For example, different species of trees often have
different LAI’ and height, while in forest patches with similar forest
species the spacing and orientation of the trees may be different.
Similarly, the land use history or resource limitations of the local
area could be important for coarse and fine-scale forest structure
information (Stephens et al., 2021). It’s worth noting that these forest
structure metrics are highly correlated across the full domain, but
their correlation within elevation bands is more mixed due to more
homogeneity in forest species, terrain, land use history, and other
factors (Figure 4). fVEG is the most important vegetation metric in our
analysis, compared with openness and LAI’, is easier to estimate, and
matches the linear relationship found in Varhola et al. (2010).
Importantly, however, Varhola et al. (2010) did not explore finer scale
forest structure information like openness and LAI’. Varhola et al.
(2010) also emphasized the inconsistencies between forest structure
metrics across studies, which hints at the value of using lidar
information to develop repeatable coarse and fine forest structure
information.
Management applications and future research directions
While simple linear models capture broad patterns of forest-snow
interactions, thresholds at which specific forest structure metrics
become important can help refine forest management decisions beyond the
current decision-making paradigm that typically does not include
benefits to snow water resources. We developed a decision support tool
for SCB, and extrapolated that to a landscape scale, by defining
qualitative thresholds in fVEG and openness that correspond to high,
moderate, and low response of SWE to potential thinning. The Sagehen
Project was a mixture of mechanical thinning and prescribed fire that
was planned to address wildlife and fire concerns and resulted in the
fragmented treatment areas shown in Figure 1. We find near zero DSnowAc
(~0%) and more absolute SWE (~50 cm) in
the low response areas of SCB, suggesting that our threshold method has
identified areas with little response to thinning or other canopy
removal. These areas account for a small portion (~7%
in 2014) of the total domain, suggesting medium and high response areas
were widespread across SCB prior to thinning from the Sagehen Project.
The high response areas had a delta SWE of around -25% and lower
absolute SWE of ~30 cm. Only 29% of the treated areas
as part of the Sagehen Project thinning were high response areas
(pre-thinning), while 27% were moderate response. While the resilience
of snowmelt-derived water resources was not an explicit priority for the
Sagehen Project, it illustrates the potential co-benefits of
simultaneously managing forest fire reduction, wildlife habitat, and
snow retention . Our decision support tool offers both a retrospective
means to assess restoration effectiveness (as done in the Sagehen
Project area), as well as a tool for proactive planning efforts.
One of the key advantages of our results for decision support purposes
is that it can be extrapolated to larger domains of similar climate and
forest conditions or re-developed in new areas with different snow
conditions (e.g. colder and windier) that have existing snow-on and
snow-off lidar datasets. To demonstrate the scalability of our results
for decision support, we extended our low, medium, and high response
mapping to an 800 km2 landscape-scale area with
planned forest restoration by the USFS. Lidar analyses showed that
treated areas had a higher relative proportion of low response zones
compared to areas that had not yet been treated, suggesting that
treatment planning was optimizing areas for increased snow retention.
The increase in the treatments in high response zones after 2014
suggests that this type of lidar-based analysis could have been used
proactively to plan treatments for maximum snow effects. Areas near SCB
that have burned in recent years (Figure 11b), especially those that
have converted to shrub cover, show a much higher proportion of low
response areas compared to surrounding unburned and un-thinned
forestland. This difference suggests that the fires were effective at
changing forest structure in a manner that would increase snow
accumulation and retention, but that land cover conversion from forest
to shrub may have impacts not estimated with our decision support tool.
While wildfires may increase ablation due to black carbon reducing
albedo several years post-fire , other studies have shown that
reductions in canopy cover from fire can provide net benefits to
snowpack and streamflow in the Sierra Nevada . Our method does not
consider the effects of fire or other processes that disturb the snow
albedo or energy budgets beyond physical changes in forest structure.
The simplicity of our methods suggests that replicating our method in
different forest types, different forest restoration treatments, and
across a range of climates with lidar data coverage could yield both
process insight and new decision support tools.