Conflict of Interest Disclosure
We have no conflict of interest to disclose.
Abstract
Wildfires and heatwaves have recently affected the hydrological system
in unprecedented ways due to climate change. In cold regions, these
extremes cause rapid reductions in snow and ice albedo due to soot
deposition and unseasonal melt. Snow and ice albedo dynamics control net
shortwave radiation and the available energy for melt and runoff
generation. Many albedo algorithms in hydrological models cannot
accurately simulate albedo dynamics because they were developed or
parameterised based on historical observations. Remotely sensed albedo
data assimilation (DA) can potentially improve model performance by
updating modelled albedo with observations. This study seeks to diagnose
the effects of remotely sensed snow and ice albedo DA on the prediction
of streamflow from glacierized basins during wildfires and heatwaves.
Sentinel-2 20-m albedo estimates were assimilated into a
glacio-hydrological model created using the Cold Regions Hydrological
Modelling Platform (CRHM) in two Canadian Rockies glacierized basins,
Athabasca Glacier Research Basin (AGRB) and Peyto Glacier Research Basin
(PGRB). The study was conducted in 2018 (wildfires), 2019 (soot/algae),
2020 (normal), and 2021 (heatwaves). DA was employed to assimilate
albedo into CRHM to simulate streamflow and was compared to a control
run (CTRL) using off-the-shelf albedo parameters. Albedo DA benefited
streamflow predictions during wildfires for both basins, with a KGE
coefficient improvement of 0.18 and 0.20 in AGRB and PGRB, respectively.
Four-year DA streamflow predictions were superior to CTRL in PGRB, but
DA was slightly better in AGRB. DA was not beneficial to streamflow
predictions during heatwaves. These results show that albedo DA can
reveal otherwise unknown albedo and snowpack dynamics occurring in
remote glacier accumulation zones that are not well simulated by model
predictions alone. These findings corroborate the power of observational
tools to incorporate near real-time information into hydrological models
to better inform water managers of the streamflow response to wildfires
and heatwaves.
Keywords: Albedo, Data Assimilation, Wildfires, Heatwaves, Cold
Regions Hydrological Model (CRHM), Streamflow Prediction, Glacier
Hydrology, Sentinel-2.
1. Introduction
In an era of global environmental change, hydrological models need to
account for processes resulting from unprecedented combinations of
forcing meteorology, state variables, and parameters. Although the
hydrological community has made advances in creating physically based
process hydrology models, many process representations are based on
historical behaviour. Recent wildfires and heatwaves worldwide,
especially in Canada (Baars et al. , 2019; Parisien et al. ,
2023), challenge the calculation of net shortwave radiation using snow
and ice albedo algorithms based on historical representations. Snow and
ice surfaces can be darkened by wildfire soot deposition. Likewise, snow
surfaces can be darkened by accelerated snowmelt caused by rapid
above-average temperature changes from heatwaves. These two conditions
strongly impact melt energy for the seasonal snowpack, perennial
snowfields and mountain glaciers, resulting in faster melt than might be
estimated without consideration of rapid surface darkening.
Shortwave (SW) radiation input into glacierized basins is often the most
important source of available energy for melt and runoff generation.
Albedo, therefore, controls the amount of SW radiation entering snow and
ice surfaces and the availability of melt energy. The mechanism
controlling snow and ice albedo decrease due to wildfire soot deposition
is straightforward. It depends on the amount of soot deposited over a
surface and whether or not that soot would be washout by melt or be
further developed by algae growth (Aubry-Wake et al. , 2022a;
Bertoncini et al. , 2022; Esser et al., accepted manuscript).
Heatwaves, in contrast, have a more intricate effect on snow and ice
albedo. Rapid above-temperature changes can cause accelerated snowmelt
(Koboltschnig et al. , 2009; Box et al. , 2022) and
consequently earlier exposure of firn and ice; however, these same high
temperatures will most likely not further decrease the albedo of firn
and ice in nature. It is unknown whether or not current hydrological
model albedo algorithms are able to account for such different interplay
of environmental conditions in a nonstationary changing climate, given
they were developed and parameterised based on historical observations.
Up-to-date albedo observations are, therefore, necessary to update
hydrological model albedo algorithms. The availability of near
real-time, high-resolution satellite data with shorter revisit times
combined with the advancement of albedo retrieval algorithms has
provided superior quality albedo data for effective assimilation into
hydrological models.
Glacio-hydrological models are a set of numerical representations of
hydrological processes that together culminate in the ability to predict
surface water and energy budget terms, the states of soil moisture,
groundwater storage, snowpacks, and glacier mass balance and fluxes of
evaporation, sublimation, and streamflow. These models are forced with
meteorological variables and parameterized with environmental
information, and they can be used to simulate hydrologically-relevant
variables such as snow water equivalent (SWE) (Wrzesien et al. ,
2017; Marsh et al. , 2020) to unobserved locations or to diagnose
previous flood (Hamlet and Lettenmaier, 2007; Pomeroy et al. ,
2016) and drought (Fang and Pomeroy, 2007; Mishra and Singh, 2011)
events. Numerical weather precipitation models can force hydrological
models for short-term flood forecasting to prepare riverine communities
for flooding (Alfieri et al. , 2013; Thieken et al. , 2023).
Finally, climate projections coupled with hydrological models can guide
governments and environmental planners regarding a region’s future state
of water resources (Milly et al. , 2002; Blöschl et al. ,
2019). All these applications make hydrological modelling a crucial tool
to learn from the past and prepare for short- and long-term changes in
the hydrological cycle to avoid the large cost of floods and droughts.
The global costs of floods and drought-related (plus heatwaves and
wildfires) extreme events attributed to climate change amounts to US$
127 billion per year between 2000 and 2019 (Newman and Noy, 2023).
There are three main types of hydrological models depending on the
manner in which how hydrological processes are represented. These range
from empirical, semi-empirical, to physically based hydrological models
(Beven, 2012). Ideally models should represent hydrological processes as
physically-based as possible (Paniconi and Putti, 2015); however, often
inefficiencies in model physical processes development are masked by
heavy calibration (Menard et al. , 2021). The physically based
effort is a pledge to make hydrological models more robust to
unprecedented environmental conditions (Kreibich et al. , 2022).
Models that heavily rely on empiricism to represent hydrological
processes are more likely to be deemed unsuccessful under conditions
that have not been observed in the past. For instance, the uncertainty
in end-of-century mean flows can be over 40% due to the choice of
hydrological models of various degrees of physical process
representation (Krysanova et al. , 2017). Moreover, these
unprecedented environmental conditions will likely be more common under
climate change-induced modifications in the overall hydrological system
(Blöschl et al. , 2019; Queen et al. , 2021). Therefore, it
is crucial that process representation in hydrological models be as
physically based as possible for riverine communities to be prepared for
upcoming extreme hydrological events. However, developing robust
physical process representation is time-consuming. In the meantime,
other techniques should be explored to account for inherent process
representation inefficiencies that are still represented
semi-empirically in physically based models (Beven, 2012).
One way to correct for the lack of physical representation in
hydrological modelling is through DA. In simple terms, DA tries to
create an optimal estimate (\(\hat{x}\)) of a true real-world state
(\(x\)) by combining a modelled state (\(m\)) and a corresponding
observation (\(o\)). The optimal state \(\hat{x}\) is a weighted sum of\(m\) and \(o\). The weights are defined by the uncertainty in \(m\) and\(o\), favouring the less uncertain estimate of \(\hat{x}\). The
observation uncertainty, or measurement error, is usually defined a
priori based on the literature or ground-truthing. On the other hand,
the modelling error can be determined using many available techniques,
which are usually based on some sort of Monte Carlo simulations that can
capture the spread in multiple model trajectories (Reichle, 2008). The
most commonly used technique in hydrology and snow modelling is the
Ensemble Kalman Filter (EnKF) (Andreadis and Lettenmaier, 2006; Clarket al. , 2006, 2008; Slater and Clark, 2006; Huang et al. ,
2017; Lv and Pomeroy, 2020), which is a less computationally expensive
version of the previously developed Extended Kalman Filter (EKF)
(Reichle et al. , 2002). The EnKF method calculates the modelling
error based on an ensemble of simulated model states forced by perturbed
variables. The simulations are carried out until the most recent
available observation. Then, the Kalman gain is calculated, allowing the
optimal state estimate to be calculated and replaced in the model for
future simulations (Evensen, 1994). The EnKF is advantageous against
other smoothing DA techniques because it can be implemented in a
forecasting mode by calculating modelling error using only one available
observation (Reichle, 2008). This EnKF characteristic is suitable for
improving streamflow forecasting systems (Huang et al. , 2017).
Among many snow and hydrology variables, remotely sensed albedo DA has
been utilized in large-scale land surface models (LSMs) to estimate SWE
(Dumont et al. , 2012; Malik et al. , 2012; Wang et
al. , 2015), but not yet in a full hydrological model capable of
predicting streamflow.
Some hydrological models, especially those dedicated to the simulation
of snow and ice processes important to cold regions, have a dynamic
albedo simulation algorithm. Accurate simulation of albedo is critical
to define the net shortwave radiation of snowpacks and glaciers, as
shortwave radiation is the primary source of available energy for melt.
Albedo algorithms intend to simulate temporal changes in surface albedo
within a spatial modelling unit arising from soil moisture variations,
vegetation phenology, addition of fresh snow cover, snow depletion, and
firn and glacier ice exposure. In the case of partially or completely
covered snowpacks, the albedo varies due to solar angle diurnal and
seasonal variations, snow grain size (Marks and Dozier, 1992), the
amount of snow-free surfaces (Pomeroy et al. , 1998), and the
amount of light-absorbing particles (LAPs) in snow (Warren and Wiscombe,
1980). Empirical snow albedo estimation algorithms have been developed
and applied satisfactorily in the past (Gray and Landine, 1987;
Verseghy, 2012); however, the conditions to which they were developed
have changed drastically with climate change. In the case of bare ground
and short vegetation, these algorithms usually apply a constant known
albedo for fresh snow and a decay function until it reaches a depth in
which the snow-free albedo starts to contribute, and ultimately, a
constant snow-free albedo is used. The albedo is reset when fresh snow
accumulates on the ground, and the decay function starts again.
Physically based radiative transfer algorithms have been developed to
simulate snow and ice albedo (Wiscombe and Warren, 1980; Gardner and
Sharp, 2010). Some of these physically based algorithms can also account
for the introduction of LAPs into the snowpack (Zhang et al. ,
2017; McKenzie Skiles et al. , 2018). Although these physically
based algorithms should be more robust in their ability to estimate
albedo in a changing climate with more wildfires and heatwaves, they are
usually complex and, therefore, rarely implemented in hydrological
models (Pietroniro et al. , 2007; Bergström and Lindström, 2015;
Hamman et al. , 2018; Pomeroy et al. , 2022), especially
algorithms capable of accounting for LAPs deposition. Studies usually
determine the LAPs radiate forcing and its respective melt, but are
unable to directly couple them into a hydrological model (Flanneret al. , 2007; Zhang et al. , 2017; McKenzie Skiles et
al. , 2018; Magalhães et al. , 2019), unless the model is directly
forced by albedo observations (Aubry-Wake et al. , 2022a).
Albedo can also be estimated using remotely sensed imagery. Spectral
albedo can be simply calculated by dividing the reflected spectral
radiance by the incoming solar radiation at the surface for a given
region of the solar electromagnetic spectrum. Besides hyperspectral
sensors, most remote sensing systems only have a few narrow spectral
bands, and a conversion to the full broadband solar spectrum is
necessary to calculate albedo. This conversion is usually done using
narrow-to-broadband equations developed for the most common sensors
using field spectroscopy libraries or radiative transfer simulations
(Liang, 2000; Greuell et al. , 2002; Li et al. , 2018). In
addition, because most high-resolution multi-spectral remote sensing
systems operate at nadir viewing angles, they cannot observe albedo
variations due to different sensor-solar geometries. One way to overcome
that is to use coarse resolution systems with off-nadir capabilities,
such as the Visible Infrared Imaging Radiometer Suite (VIIRS) or the
Moderate Resolution Imaging Spectroradiometer (MODIS), to calculate a
Bidirectional Reflectance Distribution Function (BRDF) (Roujean et
al. , 1992; Wanner et al. , 1995; Li et al. , 2001). This
BRDF information can be downscaled to Landsat-era (30-m) and Sentinel-2
(20-m) resolutions in soil and vegetation surfaces (Shuai et al. ,
2011). These methodologies have been successfully applied for soil and
vegetation surfaces, but they were suboptimal on snow and ice surfaces
due to sensor saturation prior to the launch of Landsat-8. With the
advancement in sensor radiometric technology onboard Landsat-8 and
Sentinel-2 platforms, surface reflectance can be estimated over snow and
ice at high-resolution and, in conjunction with BRDFs, calculate albedo
over snow and ice (Wang et al. , 2016; Li et al. , 2018;
Bertoncini et al. , 2022). The high revisit time of Sentinel-2
platforms has allowed such technologies to capture rapid snow and ice
albedo changes caused by wildfire soot deposition and its associated SW
radiative forcing (Bertoncini et al. , 2022).
The degree to which streamflow predictions during melt of snow and ice
can be improved by high-resolution remotely sensed albedo DA into
hydrological models has not been assessed yet. Given the importance of
simulating accurate current and future streamflows and the increasing
trends of unprecedented wildfire and heatwaves (Jolly et al. ,
2015; Kirchmeier-Young et al. , 2019; Al-Yaari et al. ,
2023; Parisien et al. , 2023), accounting for these unrepresented
processes in glacio-hydrological models can be crucial. Because current
hydrological model’s albedo algorithms are not able to account for rapid
and unseasonal snow and ice albedo changes (Pietroniro et al. ,
2007; Bergström and Lindström, 2015; Hamman et al. , 2018; Pomeroyet al. , 2022; Wheater et al. , 2022), albedo DA provides a
path forward in improving this process’s representation. Modular
physically based models such as CRHM (Pomeroy et al. , 2022)
provide a suitable platform to test albedo DA in cold regions. Although
coarse-resolution remotely sensed snow and ice albedo DA into LSMs has
been performed to estimate SWE in the past (Dumont et al. , 2012;
Malik et al. , 2012; Wang et al. , 2015), not until recently
that high-resolution remotely sensed snow and ice albedo estimates
became reliable and frequent enough (Li et al. , 2018; Bertonciniet al. , 2022) to have an impact on hydrological model streamflow
predictions. Therefore, no studies have assessed the impact of remotely
sensed high-resolution albedo DA into a cold regions hydrological model
on streamflow predictions during extreme wildfire and heatwave
conditions, especially in glacierized mountain headwater basins.
The purpose of this chapter is to diagnose the effects of remotely
sensed high-resolution snow and ice albedo data assimilation on the
prediction of streamflow of high mountain glacierized basins during
wildfires and heatwaves. The specific objectives are (i) to develop and
evaluate a cloud-computing remotely sensed snow and ice albedo retrieval
framework, (ii) to develop a framework for remotely sensed albedo DA
into a physically-based cold regions hydrological model, and (iii) to
assess the impact of albedo DA on streamflow prediction of glacierized
basins during wildfire and heatwave conditions. The albedo DA framework
was developed and tested in the Athabasca Glacier and Peyto Glacier
research basins in the Canadian Rockies during contrasting environmental
conditions that included wildfires and heatwaves between the 2018 and
2021 water years (WYs). The framework was developed based on 20-m
Sentinel-2 imagery albedo estimates that were assimilated into CRHM to
assess the impact of wildfires and heatwaves on streamflow predictions.
2. Material and Methods
2.1. Study Area
Two glacierized basins in the Canadian Rockies, Alberta, were used as
study areas for this research: Athabasca Glacier and Peyto Glacier
research basins (Figure 1). AGRB is in Jasper National Park and is part
of the Global Water Futures Observatory (GWFO) and operated by the
Centre for Hydrology, University of Saskatchewan. AGBR is a glacier
outlet of the Columbia Icefield, the largest icefield in the Canadian
Rockies and a triple continental drainage divide between the Mackenzie,
Nelson, and Columbia river basins which flow into the Arctic, Atlantic,
and Pacific oceans, respectively. AGBR has an area of 29.3
km2 and sits between 1926 and 3459 m of elevation (as
of 2011), with 58% of its area covered by glacier ice (as of 2016)
(Pradhananga and Pomeroy, 2022). AGBR is equipped with two automatic
weather stations (AWS), Athabasca Ice and Athabasca Moraine, and one
streamflow gauge. PGRB drains into the Nelson river basin. PGRB has an
area of 22.4 km2 and sits between 1907 and 3152 m (as
of 2014), with 44% of its area covered by glacier ice (as of 2016)
(Pradhananga and Pomeroy, 2022). PGRB has one AWS (Peyto Main) and a
streamflow gauge and has been the subject of intense scientific studies
since the 1960s. For more information about these research basins and
their instruments the readers are referred to Pradhananga et al.(2021) and Pradhananga and Pomeroy (2022).