Figure 8: PGRB spring and summer DA and CTRL albedos at the snow and
glacier ice HRUs highlighted in Figure 1. The 2018 and 2021 years are
shown to represent wildfire and heatwave conditions, respectively.
4. Discussions
4.1. Albedo DA During Wildfire
and Heatwave Conditions
The results presented in the previous sections have demonstrated that
albedo DA can improve streamflow simulations during wildfires but not
during heatwaves. The streamflow improvement response to albedo DA in
the soot-feeding algae year was only considerable in PGRB. These results
reveal somewhat contrasting processes happening in different zones of
these basins. Over glacier ice, DA decreased albedo considerably for
AGRB due to wildfire soot deposition, a process that was expected and
confirmed in previous studies (Aubry-Wake et al. , 2022a;
Bertoncini et al. , 2022). In PGRB, the decrease in ice albedo due
to DA was not as pronounced because of prolonged spring and summer
snowcover over ice. SWE is another state that is updated proportionally
to albedo. Because these states are the mean of 20 ensembles, the
likelihood of all ensembles converging in the absence of a snowpack
becomes lower when several ensembles present elevated SWE values.
Figures 5 and 6 show that the SWE ensemble spread in PGRB was wider than
in AGRB, contributing to a shorter period of exposed ice in PGRB. This
mechanism could have been caused by deeper snowpacks observed in
terminal sections of PGRB and more frequent spring and summer snowfall
events. The effect of prolonged snowcover when compared to control
simulations in snow DA has been reported before, usually leading to snow
depletion simulations closer to observations (Smyth et al. , 2020;
Alonso-González et al. , 2022). It is worth noting that once the
snow is depleted and firn and ice are exposed, temperature-driven albedo
decrease ceases. This mechanism should be captured by the albedo decay
algorithm that uses constant albedo values for exposed firn and ice. The
latter can potentially explain why streamflow predictions were not
sensitive to albedo DA during the heatwave year.
Unlike ice, snow has a different response to albedo DA. Albedo DA has
shown to be larger than modelled by CTRL in AGRB high-elevation
snow-dominated regions. The introduction of remotely sensed albedo
through DA has revealed that snow was not completely melted in the AGRB
high-elevation HRU examples displayed in Figure 7, i.e., albedo did not
reach the 0.55 firn value. The low CTRL snow albedos can be a limitation
of albedo algorithms based on decay functions, such as the one used
hereby, which were developed for seasonal snowpacks at much lower
elevations. A comparison of three empirical models with a full
physically based model (closer to observations) has shown that empirical
decay albedo models, indeed, underestimate snow albedo (Gardner and
Sharp, 2010). On the other hand, DA snow albedo is often below CTRL in
PGRB, but rarely reaches the firn value of 0.55 (Figure 8). This result
suggests that snow-dominated PGRB surfaces would have a lower DA albedo
than CTRL, since they are more heterogeneous due to greater firn and ice
exposure than in AGRB. CTRL seems to miss processes well described at
lower elevations but not at glacier accumulation zones in both basins.
This finding calls for a better representation of glaciological albedo
processes capable of accounting for the peculiarities of localized
effects (Marshall and Miller, 2020).
This study tested two main assumptions by introducing remotely sensed
albedo in a cold regions DA framework. First, soot deposition would
decrease snow and ice albedo during extreme wildfire activity, such as
in 2018. This assumption was confirmed by previous studies (Aubry-Wakeet al. , 2022a; Bertoncini et al. , 2022) and hereby for ice
in AGRB and snow in PGRB. A greater albedo in the AGRB snow HRU example
can be explained by the larger elevation range observed between snow and
ice regions in AGRB (Pradhananga and Pomeroy, 2022). This larger
elevation range creates a scenario in which snow-covered plateaux are at
elevations more prone to new snowfall, suppressing the effect of soot
deposition over snow at a faster pace. Rapid recovery from soot
deposition over high-elevation snow in AGRB has been previously
described in Bertoncini et al. (2022). Second, heatwaves would
accelerate melt and expose ice or firn unseasonally. This process does
not seem to substantially affect AGRB ice since exposure in 2021 is
similar to other years. However, albedo is lower for snow during
heatwaves when compared to wildfire years. Substantial snowmelt and
decreased albedo in glacier snow due to extended periods of
above-average temperature have been reported before in Greenland (Boxet al. , 2022) and the Austrian Alps (Koboltschnig et al. ,
2009). This mechanism suggests that heatwaves have a greater influence
on high-elevation snow-covered plateaux albedo decreases than does
wildfire soot in AGRB. On the other hand, there is a larger sensitivity
to wildfires than heatwaves for PGRB snow, perhaps because of more firn
and ice exposure than in AGRB. PGRB ice shows greater albedo decrease
during heatwaves due to longer ice exposure. In addition, the albedo
algorithm is expected to be more robust during heatwaves than wildfires
because it considers daily mean temperatures, limiting most albedo DA
benefits to wildfire years.
4.2. Implications for Streamflow
Prediction in a Changing Climate
Unprecedented wildfires and heatwaves are expected to increase with
climate change (Jolly et al. , 2015; Kirchmeier-Young et
al. , 2019; Al-Yaari et al. , 2023; Parisien et al. , 2023)
and, hence, are likely to affect glacier contribution to streamflow.
Although this is a somewhat intuitive assumption, this study has
demonstrated inter- and intra-basin peculiarities in how the effects of
wildfires and heatwaves will contribute to snow and ice melt. The
findings show that wildfires and heatwaves can decrease glacier ice
albedo, but the period that ice is covered by snow is the primary
governing factor controlling ice melt. Likewise, the amount of summer
precipitation falling as snowfall is another factor governing albedo
dynamics in snow-dominated regions. This mechanism creates a scenario
where the elevation difference from the terminal glacier to
high-elevation snow plateaus dictates whether these basins would be
affected by either wildfires or heatwaves. For instance, in AGRB where
this elevation difference was higher than in PGRB, DA generated an
increase in snow albedo. This finding alone reveals something that would
not be possible using modelling alone; remotely sensed albedos were
needed in a DA modelling framework to understand this process in a
virtually inaccessible region. There lies the power of observational
tools in aiding hydrological models beyond the conditions in which they
were developed.
There are many implications of remotely sensed albedo DA for streamflow
predictions under climate change. The most important one is that this
study showed that even though DA was not beneficial for streamflow
prediction during heatwaves, it was beneficial in the overall four-year
evaluation in PGRB and favourable during wildfires and similar to CTRL
in other years in AGRB. From this perspective, remotely sensed albedo DA
is recommended under unprecedented hydrological extremes imposed by
climate change; however, caution should be taken when interpreting
albedo DA results during heatwaves. The latter calls for further
investigation into other processes that might have contributed to albedo
DA degradation of streamflow predictions under heatwaves beyond what has
been discussed here. Remotely sensed albedo DA can also better inform
hydrological modelling during wildfires and heatwaves. For instance,
events of decreased albedo can happen in the future in high-elevation
snow-dominated regions but be buffered subsequently by fresh snowfall,
which can only be confirmed with certainty via remotely sensed albedos
since precipitation measurements are usually taken at much lower
rainfall-dominated elevations. Users of hydrological predictions can
then understand de facto whether these extreme events will affect
downstream streamflows.
4.3. Uncertainty within DA
Modelling Framework
The DA modelling framework has worked satisfactorily in most simulated
years; however, a few factors contributed to the uncertainty in
modelling streamflow and other model states. First, although snow and
ice remotely sensed albedo estimates were satisfactory here and
elsewhere (Wang et al. , 2016; Li et al. , 2018; Bertonciniet al. , 2022), there is still a 5% uncertainty (\(\sigma_{o}\) =
0.046) in albedo estimation. This 5% uncertainty cannot be neglected in
glacier ice albedo values. Therefore, even if modelling uncertainty is
larger than 5%, this value would be, on average, the smallest
uncertainty possible of an optimal albedo estimate for a particular
assimilation date. Second, assimilation frequency can contribute to the
influence of DA in modelling albedo and other states. AGRB (68
estimates) had more than double the number of assimilation dates of PGRB
(33 estimates). The lower number of assimilation dates in PGRB could
have contributed to longer periods of snowcover over ice, since albedo
correction from a new remotely sensed update would take longer to occur.
Finally, uncertainty in hydrological modelling can also affect
streamflow prediction within a DA framework. This study presented DA NSE
values of 0.74 and 0.78 for AGRB and PGRB, respectively. These NSE
values are higher than the mean of maximum values (0.64) found in 20
studies that conducted uncalibrated streamflow predictions using the
CRHM model (Pomeroy et al. , 2022). Nonetheless, these NSE values
are not perfect, i.e., equal to 1, representing that there are still
uncertainties in streamflow prediction that can be attributed to model
structure, forcing errors, parameter errors, algorithm deficiencies, and
uncertainty caused by the DA EnKF implementation.
5. Conclusions
This research implemented and tested a remotely sensed albedo DA
framework to predict streamflow in two highly glacierized Canadian
Rockies’s basins during environmental conditions ranging from normal,
wildfire, and heatwave dominated. Glacier ice remotely sensed albedos
presented satisfactory evaluation results (r = 0.96, bias = 0.026, RMSE
= 0.060, and \(\sigma_{o}\) = 0.046) that were needed for assimilation.
Albedo DA improved streamflow predictions in the heavily
wildfire-impacted year of 2018 for both basins – a KGE improvement of
0.18 and 0.20 for AGRB and PGRB, respectively. DA in PGRB was beneficial
for all years but 2021. In the soot-feeding algae year, streamflow
improvement due to albedo DA was only considerable in PGRB. DA
substantially enhanced overall four-year streamflow prediction in PGRB
but just slightly in AGRB. DA’s streamflow prediction improvements were
caused by a balance between changes in albedo of high-elevation snow and
glacier ice. In AGRB, snow albedo was increased by DA due to frequent
summer snowfall events that buffered the streamflow generated from
decreased glacier ice albedo in lower elevations. In PGRB, snow albedo
was decreased by DA, especially during wildfires. However, glacier ice
DA albedo was only decreased during short periods of ice exposure caused
by a prolonged spring and summer snowpack. The latter mechanism results
from several ensembles generating elevated SWE values during spring and
summer in PGRB glacier ice. These findings reveal that wildfires and
heatwaves are capable of decreasing glacier ice albedo, but the
resultant melt contribution to streamflow within a DA framework will
depend on snowpack albedo and SWE dynamics.
The cloud-computing remotely sensed snow and ice albedo retrieval
framework developed in this study could generate results with comparable
accuracy to previous studies, while providing global reproducibility at
high spatial and temporal resolutions. Before this study, albedo in
snow-dominated glacier accumulation zones was based solely on albedo
modelling developed at relatively lower elevations. This albedo
representation could not account for the rapid recovery of albedo with
fresh snowfall during wildfire and heatwave seasons. This finding was
only possible utilizing high-resolution remotely sensed albedo estimates
that could reach virtually inaccessible regions. The assimilation of
these remotely sensed albedo estimates into the physically based CRHM
model improved streamflow predictions for most of the analyzed years.
Moreover, using albedo DA revealed contrasting processes happening in
poorly observed glacier zones that resulted in different streamflow
responses to wildfires and heatwaves. Considering that the environmental
conditions observed during the study are expected to increase in a
future of climate change, it can be advantageous to use remotely sensed
high-resolution snow and ice albedo DA continuously for better
streamflow predictions in glacierized basins during wildfires and
heatwaves. This study’s findings also indicate that the response of
glacierized basin streamflow to wildfires and heatwaves is not always as
expected due to the interplay of different factors such as fresh
snowfall, soot deposition, and unseasonal melt with the albedo
algorithm. Using observational tools such as DA can help narrow water
managers’ uncertainty when making decisions based on hydrological
predictions under a warmer and more wildfire prone future.
Acknowledgements
We wish to thank NASA for the MODIS data and the European Space Agency
(ESA) for the Sentinel-2 and ERA5-Land data used to estimate remotely
sensed high-resolution albedos. We also acknowledge the support of the
GEE platform for hosting the above data and providing cloud-computing
resources, and the developers of the 6S atmospheric correction model and
its Python implementation (Py6S). The help from Xing Fang to QC
meteorological forcings and to set up the CRHM model is much
appreciated. We thank Caroline Aubry-Wake and Dhiraj Pradhananga for the
discussions about glacio-hydrological modelling and the many Centre for
Hydrology field technicians who contributed to maintaining the AWS. We
also would like to acknowledge the permission of Parks Canada for
allowing our research to take place in the Banff and Jasper National
Parks and Pursuit for assistance in logistics on Athabasca Glacier.