1. Introduction
Mountain snowpack plays a critical role in the hydroclimate of the U.S. West Coast. With precipitation falling predominantly during winter, freshwater is largely stored in mountain snowpack along the Cascade and Sierra Nevada mountain ranges between November and March. With warmer temperature and increasing sunlight, snowpack begins to melt around April, providing snowmelt water for runoff that supports agriculture, fish, hydropower generation, and a host of other human activities during the otherwise dry summer season with little precipitation. Understanding the climatic control on snowpack in the U.S. West Coast is important for improving predictions and management of water resources.
Sea surface temperature (SST) is an important driver of hydroclimatic variability of the western United States. SST anomalies in the tropical Pacific Ocean associated with El Niño–Southern Oscillation (ENSO) are well known to influence temperature and precipitation in the western United States and many other regions around the world (Redmond and Koch 1991; Dettinger et al. 1998). Through their impacts on convection and diabatic heating, SST anomalies in the tropical Pacific Ocean induce Rossby waves that result in the Pacific–North American (PNA) teleconnection (Barnston and Livezey 1987). The PNA pattern influences precipitation and temperature in the western United States on a seasonal-to-interannual time scale. On interannual-to-decadal time scale, SST anomalies associated with North Pacific Oscillation (NPO) also have an influence on the hydroclimate of the western United States (Linkin and Nigam 2008). SST forcing associated with ENSO and NPO accounts for about 20% of winter precipitation variance in the U.S. West Coast (Dong et al. 2018).
As narrow, elongated pathways of concentrated moisture transport across the subtropics, atmospheric rivers (ARs) connect SST forcing in the tropical Pacific Ocean with extreme precipitation in the western United States (Dong et al. 2018; Gershunov et al. 2017; Gershunov and Cayan 2003; Newell et al. 1992; Ralph et al. 2019; Waliser and Guan 2017; Dettinger et al. 1998; Ralph and Dettinger 2011). ARs play a central role in the hydroclimate of the U.S. West Coast. Previous studies have revealed a close connection between ARs and winter precipitation on the West Coast, especially heavy precipitation (Guan et al. 2010; Chen et al. 2018) and flooding (Ralph et al. 2006; Neiman et al. 2011). ARs also play an important role in the snow process in the West Coast mountains (Chen et al. 2019b; Guan et al. 2016; Goldenson et al. 2018). By intensifying snow ablation through rain-on-snow events, ARs nearly double the runoff response to precipitation compared to non-AR events (Chen et al. 2019b). The positive correlation between rain-on-snow events in the western United States and SST in the tropical Pacific Ocean identified by McCabe et al. (2007) supports the influence of tropical Pacific SST forcing on ARs. More generally, El Niño tends to increase snowpack in the southwestern United States but reduce snowpack in the northwestern United States (Clark et al. 2001; Jin et al. 2006).
Besides the remote SST in the tropical Pacific Ocean, SST in the northeastern Pacific Ocean adjacent to the U.S. West Coast may also have an effect on the hydroclimate of the region (Potito et al. 2006; Sagarika et al. 2016). Located on the pathway of the Pacific storm tracks, SST in the northeastern Pacific (or local SST hereafter) can influence the atmospheric temperature and moisture of the winter storms, with subsequent impact on the precipitation produced by the storms and the surface air temperature in the mountains where precipitation interacts with snowpack (Persson et al. 2005; Chen and Leung 2020). It is important to note that interannual variability of SST in the tropical Pacific associated with ENSO exerts a strong influence on the local SST near the U.S. West Coast. Such influence is associated with the generation of oceanic equatorial Kelvin waves (McPhaden and Yu 1999) that transport the signal of the thermocline in the tropical Pacific to the coast and influence coastal upwelling, as well as ENSO’s impact on the Aleutian low (Alexander et al. 2002) that influences the alongshore winds and coastal upwelling. Based on these mechanisms, larger warm local SST anomalies are generally associated with strong El Niño events. However, the relationship between ENSO and local SST is more complicated, as exemplified by the extreme and persistent local SST warming during the winter of 2014/15 when the tropical Pacific SST anomalies were weak. Capotondi et al. (2019) noted that even weak tropical SST anomalies in the right location can significantly influence SST in the northeastern Pacific.
Regardless of whether or how local SST may be related to remote SST anomalies and large-scale circulation, local SST may influence the hydroclimate of the U.S. West Coast through its effect on evaporation and sensible heat flux that influence moisture transport and atmospheric instability, and hence precipitation (Chen and Leung 2020). The connections between local SST and temperature, precipitation, and snowpack in the U.S. West Coast have been largely unexplored except in the context of modeling. Mejia et al. (2018) found systematic warm SST biases offshore of California and the Baja California Peninsula region in the multimodel ensembles of the Coupled Model Intercomparison Project phase 3 (CMIP3) and phase 5 (CMIP5). They identified a statistically significant correlation between the SST biases with precipitation biases in the western United States and confirmed the propagation of SST biases to the western United States using regional climate model experiments. They found that SST biases in the aforementioned local ocean regions can explain up to 80% of the wet precipitation biases over land. Caldwell et al. (2009) also found that warm SST biases inherited from global climate simulations are partly responsible for the overprediction of California precipitation in their regional climate simulations. These studies provide support for the impacts of local SST on precipitation in the U.S. West Coast, but we still lack an understanding of the spatiotemporal characteristics of the hydroclimate response to local SST, including changes in the mean, variability, and extreme due to the mean and variability of the local SST forcing.
In this study, regional climate simulations and statistical models developed based on neural networks are employed to investigate the impact of local SST on the hydroclimate of the western United States, focusing particularly on mountain snowpack and its meteorological drivers. We address the following three questions: 1) How does the western U.S. mountain snowpack respond to the local SST forcing? 2) What are the relative roles of precipitation and temperature in these responses? 3) What are the impacts of local SST on the moisture transport by ARs and the subsequent influence on mountain snowpack? Using both physical and statistical models, we aim to construct a complete picture of how local SST affects meteorological conditions and mountain snowpack.
This manuscript is organized as follows: section 2 describes the two simulations and the statistical models used in this analysis; results are presented in section 3 and discussions in section 4. The summary and conclusions are given in section 5.
2. Data and method
The Weather Research and Forecasting (WRF) Model is used in two regional climate simulations at convection-permitting 6-km grid spacing covering the western United States and the northeastern Pacific Ocean. The simulations are configured with the Morrison double-moment cloud microphysics scheme, the Rapid Radiative Transfer Model for GCMs (RRTMG) parameterization for longwave and shortwave radiation, the Mellor–Yamada–Janjić turbulence closure scheme, and the Unified Noah land surface model. The historical simulation (HIST) is the same as that described and analyzed in Chen et al. (2019a,b), with a simulation period covering 1 October 2000–30 September 2015, with the first three months used as model spinup. The simulation was driven by lower and lateral boundary conditions from the North American Regional Reanalysis (NARR) (Mesinger et al. 2006), which has also been used with the WRF model in other studies of snowpack in the western United States (Rasmussen et al. 2011; Hughes et al. 2020; Wrzesien et al. 2017; Gutmann et al. 2012). Our evaluation of the HIST run indicates that WRF captures over 87% of the observed precipitation variability (Chen et al. 2019a). Similar to the HIST run, we also performed a simulation (fSST) in which SST was inadvertently fixed at the initial condition at 0000 UTC 1 October 2000 throughout the simulation, while the same time-varying lateral boundary conditions as HIST were used. Hence, the only difference between fSST and HIST is the use of time-invariant versus time-varying local SST. Synoptic systems entering the boundaries of the WRF domain that influence precipitation in the western United States are identical between the two simulations, but temperature and precipitation over land are modulated by the local SST and associated air–sea heat and moisture exchanges. While local SST is known to be related to SST anomalies in the tropical Pacific (i.e., remote SST), it is treated as an independent forcing variable in our numerical experiments to facilitate understanding and quantification of the local SST effect alone. More discussion on this fSST experiment is provided in section 4.
Figure 1 illustrates the simulation domain, along with the SST prescribed in the fSST simulation as the percentile of the daily SST in the HIST simulation during the analysis period (1 October 2003 and 30 September 2015). With SST fixed at the initial condition on 1 October 2000, fSST has warmer SST than the daily SST prescribed in HIST, particularly near the California coast and in the Gulf of California (Fig. 1a). The monthly cycles of the average local SST are illustrated in Fig. S1 in the online supplemental material. Overall local SST differs by −0.5 to 4 K in winter. Fixing the SST at the initial condition in fSST also resulted in the removal of SST variability at all time scales in the fSST simulation over the northeastern Pacific and the Gulf of California.
The SST condition in the fSST experiment. The WRF simulation domain is illustrated as a dashed blue box. (a) The fixed SST level as percentiles of the daily SST during 1 Oct 2003 and 30 Sep 2015. (b) The variability of the observed winter SST in the region, which becomes all zeros in the fSST experiment. The West Coast region is delineated by the bold black lines encompassing the states of Washington, Oregon, California, and Nevada. Land regions in colors indicate the grid elevation of the mountainous area where mean 1 Apr SWE is larger than 100 mm during 2003–15.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
Different experimental designs can be used to study the snowpack response to local SST forcing. Although the fSST simulation does not correspond to any realistic scenarios, this fortuitous simulation with warm local SST and absence of local SST variability still provides useful results to address our science questions. We examine the variabilities of snow water equivalent (SWE), precipitation, and temperature at interannual, monthly, and daily scales. Our analysis focuses on regions where the mean 1 April SWE is higher than 100 mm during 2003–15, which are shown along with the surface topography information in Fig. 1. These areas are further grouped into two subregions: Washington and Oregon (north) featuring the Cascade Mountains, and California and Nevada (south), featuring the Sierra Nevada range with higher surface elevation in general. These two subregions also show dipole precipitation statistics in previous studies (Wise 2010).
Since the two regional climate simulations used the same lateral boundary conditions, we focused on the changes in AR intensity rather than AR frequency. AR intensity, measured by the vertically integrated vapor transport, and its relationship with snowpack over land is examined. The occurrence and spatial coverage of ARs are taken from the AR Tracking Method Intercomparison Project (ARTMIP) (Shields et al. 2018). ARs identified based on the Gershunov algorithm and MERRA-2 reanalysis product represent a moderate-to-large number of ARs relative to other algorithms used in ARTMIP along the West Coast (Gershunov et al. 2017; Gelaro et al. 2017; Shields et al. 2018). Similar to previous studies of snow processes in the western United States (Chen et al. 2019b), this study used the AR data from the Gershunov method (Shields et al. 2018). Although AR frequency is the same in HIST and fSST, the atmospheric moisture and moisture transport of the ARs can be different due to the local SST perturbations, which may affect mountain snowpack.
Input to the neural network NNs for cold season ΔSWE estimation.
At monthly scales, a long short-term memory (LSTM) network (NNm, where “m” stands for month) is trained to predict the SWE on the first day of each winter month. Previous studies have demonstrated the capability of LSTM to construct hydrological time series such as soil moisture and streamflow (Fang and Shen 2020; Kratzert et al. 2018; Sahoo et al. 2019). The SWE of a given month is the cumulative result of the meteorological conditions up to several months earlier, making LSTM a natural fit for this task. This network takes the meteorological data of the previous 12 months and predicts the SWE of the current month. The inputs of this model are given in Table 2.
Input to the neural network NNm for monthly SWE estimation.
SWE data (winter ΔSWE and the SWE on the first day of each winter month) from the HIST simulation are used to train the two models. As evaluated in Chen et al. (2019b), the HIST simulation captures the observed SWE records with good accuracy. Thus, training the neural networks to mimic the WRF behavior ensures that the neural networks capture the snow response in the real world with reasonable accuracy. We used 64% of the SWE data to train the models, and 16% to validate the model performance. The remaining 20% serves as the testing subset. With warmer SST than HIST, the fSST simulation can generate meteorological conditions that are not in the HIST simulation. Thus, fSST can be used to verify whether the NN models can predict the snowpack responses in a different climate, and how transferable the NN models are to other mountainous regions with different climates. The trained neural networks allow us to isolate the impact of different meteorological conditions on the snowpack response. For example, by running NNs with the meteorological conditions in HIST, we estimated the historical snow accumulation during winter ΔSWE0 = NNs(Ptot, Fwet, Tp, Tdry). Replacing one of the inputs with that of the fSST simulation, such as
3. Results
In this section, we first analyze the response of snowpack to the local SST difference between fSST and HIST, followed by an analysis of the meteorological responses (precipitation, temperature, and AR) to the local SST forcing. Then we use the neural network models to decompose the contributions of each meteorological factor to snowpack response.
a. Response of snowpack to SST perturbations
Figure 2 shows the mean winter snow accumulation (“winter ΔSWE” calculated as the difference between the SWE on 1 April and 1 October) in the two experiments. Figure S2 shows the mean SWE at the beginning (1 October) and end (1 April) of winter in the two experiments. The mean SWE on 1 October (i.e., the beginning of winter) is similar between HIST and fSST, so the difference in the snow accumulation of each winter does not affect the snowpack response of the next winter. Responding to the nearly spatially uniform warmer SST (Fig. 1a), there is a notable decrease in winter ΔSWE in fSST compared to HIST in the Cascade Range in the north but an increase in the Sierra Nevada in the south. Figures 2d and 2e illustrate the difference between winter ΔSWE (fSST − HIST) as a function of surface elevation. There is a clear elevation dependence of the winter snowpack response to the SST difference between the two simulations. In both subregions, the winter ΔSWE difference is positive at higher elevations and small or negative at lower elevations, with a transition occurring at elevations between 1800 and 2000 m. To explain this elevation dependence, we analyze the difference in snow accumulation (SWE+) and snow ablation (SWE−) separately, and the results are shown in Figs. S3 and S4. For snow accumulation, Fig. S3 indicates that relative to HIST, SWE+ days in fSST happen less frequently in the north, but more frequently in the south. Decomposing the ΔSWE difference into the contributions from snow accumulation (Fig. S3) and snow ablation (Fig. S4) indicates that the winter ΔSWE decrease in the north is a combination of less accumulation and intensified ablation, while in the south, the change is dominated by increased accumulation.
Difference of winter snow accumulation in the two experiments: the mean winter ΔSWE during 2003–15 in the (a) HIST and (b) fSST experiments and (c) the difference of mean winter snow accumulation as (b) minus (a). (d),(e) The difference in winter ΔSWE in the 12 years as a function of surface elevation. Dark blue in these two panels indicates the 50% confidence interval, and light blue indicates the 90% confidence interval.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
Figure 3 shows the change of variability of SWE comparing HIST and fSST, with and without temporal SST variability, respectively. Similar to the mean winter ΔSWE change, the interannual variability of winter ΔSWE in the north also decreases in the fSST experiment, while it increases in the south (Fig. 3c). In the following analysis, we will relate the change of snowpack variabilities to the changes in the meteorological conditions.
(a),(b) Difference in the mean of SWE and meteorological conditions and the variations (c)–(e) over the whole winter and (f)–(h) over the winter months for (left) SWE; (center) precipitation; and (right) 2-m temperature. The ΔSWE is defined as the total change of winter or each month, precipitation and temperature are the mean of winter or each month. Note the unique color bar to (b). The details of these changes can be found in Figs. S4–S6. Panel (b) is based on the upper color bar, and the rest of the panels are based on the lower color bar.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
b. Relationships between meteorological and snowpack changes
The changes of precipitation and air temperature between fSST and HIST are shown in Figs. 3a and 3b while their mean and variabilities are illustrated in Figs. S5 and S6. With warmer SST, winter mean precipitation increases in fSST relative to HIST in both subregions by 10%–20%. Precipitation variability also increases, with similar relative change in both subregions (Figs. 3d,g), which is different from the change in SWE variability that features a north–south dipole pattern (Figs. 3c,f). Meanwhile, the mountain regions, especially in the north, experience warmer winter temperatures in fSST than HIST but the variability of temperatures is noticeably reduced in both regions in fSST relative to HIST (Figs. 3e,h). The larger warming in the north could be contributed by the intensified warming in the northern part relative to the southern part of the local ocean (Fig. 1 and Fig. S1). As SWE is influenced by both precipitation and temperature, the different responses in SWE variability between the north and south may be a result of a competition between the changes in variabilities of precipitation and temperature. The relative contributions of temperature and precipitation to the north–south dipole changes in SWE variability will be explored using neural networks in section 3c.
Figure 4 shows the codistribution of P and T across the snow regions, highlighting a significant shift of precipitation events toward temperature above freezing in the north in fSST compared to HIST. For the precipitation-weighted air temperature, TP for the north is 272.4 K in the HIST simulation and 273.2 K in the fSST simulation. In comparison, they are 271.1 and 271.5 K in the south, respectively. These indicate that the snowpack in the north is more vulnerable to warming than that in the south, as the HIST TP is closer to freezing in the north (272.4 K) than in the south (271.1 K). The warmer temperature in fSST is enough to increase the frequency of TP above the freezing point, resulting in more frequent rainfall at the expense of snowfall in the north. This would lead to rain-on-snow effect that increases snowmelt, consistent with the intensified ablation shown in Fig. S4, thus increasing the variability of winter ΔSWE. In the south, however, even with the warming over land in fSST, TP is still below freezing point on average, so precipitation is still more likely to be in the form of snowfall. Thus the snow accumulation increases in the south, along with a reduction in the variability of winter ΔSWE.
Codistribution of P and T across the snow regions in (a),(d) HIST and (b),(e) fSST and (c),(f) fSST − HIST. The vertical dashed lines mark the freezing temperature. The red lines in the left and center columns indicate the P-weighted temperature in the region/simulation, and the values are noted by the numbers highlighted in red in each panel. These plots are generated using data over the mountainous regions shown in Fig. 1.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
Other aspects of precipitation change that could play a role in the changes in SWE are explored in Fig. S7. With warmer SST, wet days increase by about 5%, but the average intensity of precipitation increases by about 10%. This indicates that with the same lateral boundary conditions, warmer local SST has a larger impact on precipitation intensity than frequency. Figure 5 shows the IVT change of AR events in HIST and fSST, and the detailed analysis of the AR response is presented in Chen and Leung (2020). The northern region experiences stronger IVT than the south during ARs. In response to the warmer SST, the IVT of AR events increases across the western mountain regions. However, the northern subregion experiences a higher increase, consistent with the larger increase in surface temperature (Fig. 3b). Given ARs lead to significant snowmelt (Chen et al. 2019a), the increase of AR intensity likely contributes to the snowpack reduction in the north.
Change of atmospheric rivers (ARs) column-integrated moisture transport (cIVT) in the study region. The mean cIVT during the AR days in the (a) HIST and (b) fSST simulations and (c) the relative change of mean cIVT. Note that the AR days are identical in HIST and fSST, which correspond to the AR days detected by the Gershunov algorithm in ARTMIP.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
As revealed by Fig. 2, snowpack changes due to local SST warming shows an elevation dependence. Figure 6 visualizes the snow accumulation/ablation grouped by surface elevation band (in 200-m increment). It also shows the accumulation/ablation amount on AR days (Figs. 6c,d) and extreme AR days (Figs. 6e,f). ARs mostly influence snow ablation in areas at low to mid-elevation (<2500 m). Also, ARs are associated with both snow accumulation and ablation in the north, but in the south, ARs are more frequently related to snow accumulation, especially at higher elevations. As the warmer SST in fSST increases AR intensities (Fig. 5), snow ablation is substantially higher in fSST than HIST (Fig. 4c). Despite the less significant change of AR intensities in the south, some differences in ablation are notable in the 1500–2000-m elevation range. In both regions, differences in snow accumulation associated with differences in AR intensities are relatively small compared to the differences in snow ablation at elevations below 2500 m. Above this elevation in the Sierra Nevada, ARs slightly increase snow accumulation due to the increase in precipitation, but its overall contribution to snow accumulation is small.
Difference of snow accumulation/ablation as a function of surface elevation. (a),(b) The distribution of total snow accumulation and ablation during winter as a function of surface elevation, with the contribution from ARs marked in purple and blue. (c),(d) As in (a) and (b), but for the total accumulation/ablation caused by ARs. (e),(f) As in (a) and (b), but for the contributions from the intense ARs (i.e., IVT > local 99% percentile). In all the panels, the solid lines are the 50% percentile values from the HIST simulation, and the dashed lines are the 50% percentile values from the fSST simulation. The colored shading and transparent shading bounded by the two black lines show the 25%–75% percentile range of snow accumulation or ablation. The highest elevation in the PNW is less than 2500 m in the WRF Model.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
In summary, the contrasting responses of snowpack to the warmer local SST in fSST in the northern and southern regions are a result of the competition of snow accumulation and ablation changes. Warmer local SST increases surface air temperature and AR intensity more in the north than the south. These result in larger impacts on snow ablation in the north than the south. However, the lower elevation of the Cascades compared to the Sierra Nevada also plays an important role as temperature in the Cascades is closer to freezing than the Sierra Nevada, which has higher elevation. Hence warming has larger effects on snow ablation in the north than the south, where snowpack changes are dominated by snow accumulation due to precipitation increase. The elevation effects are amplified by AR activities as the increase in snow ablation by ARs is larger at the same elevation range between 1000 and 2400 m in the Cascades and Sierra Nevada (Figs. 6c,d).
c. Decomposition of the snowpack response
Given both P and T are critical to snow accumulation and ablation, they may be used to explain the difference of winter snowpack between the fSST and HIST experiments. ARs are characterized by heavier precipitation and warmer temperature, so they are not considered as independent factors in the neural network models. Here two models are established to relate these meteorological forcings to snowpack response at different time scales.
Figure 7 evaluates the NNs model performance in predicting the winter ΔSWE based on winter-aggregated statistics of meteorological conditions. Figure 7a compares the NNs estimated ΔSWE to the WRF simulated ΔSWE, with 93% of the spatial–temporal variation of winter ΔSWE captured by the NNs model in the training period. This model also explains 89% of the variations of winter ΔSWE in the validation datasets that are excluded in the model training. The trained model is then used to predict the winter ΔSWE in the fSST experiment. The changes of winter ΔSWE from the NNs model and the WRF simulation are compared in Figs. 7c–f. The NNs model skillfully predicts the north–south contrast in the changes of the mean and variability of winter ΔSWE. The NNs model even captures the contrasting changes on the windward and lee side of the northern Cascades, but the changes in the lower-elevation southern Cascades are less well predicted. The overall similarities between the spatial patterns of mean and variability changes indicate that NNs correctly captures the relationship between winter meteorological conditions and snowpack response to perturbations, making it a useful tool for decomposing the impact of individual meteorological factors.
Validation of NNs on the estimation of winter total snow accumulation. (a),(b) The winter snow accumulation from WRF simulation (x axis) and NNs (y axis) in the training the validation subsets, respectively. Note that both training and validation subsets are taken from the HIST simulation. (c) The WRF simulated relative change of mean snow accumulation and (d) the NNs estimated change. (e),(f) As in (c) and (d), but showing the variability of winter snow accumulation.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
With this simple model, we perform four experiments to reveal the contributions of each input variable of the NNs model (Table 1) to the difference in winter ΔSWE between fSST and HIST. The experiments are described in the method section, with results shown in Fig. 8. Changes in the four input variables are illustrated in Fig. S8. As shown in Fig. 8a, an increase in P contributes to an increase in ΔSWE, while warmer TP causes a reduction of ΔSWE. In comparison, the impacts of wet days and the temperature of nonprecipitating days are minimal. The TP and its change have a large elevation dependence following a rate similar to the moist adiabatic lapse rate (−3.5°C km−1 in HIST and −3.8°C km−1 in fSST) (Fig. S8). At lower elevation with temperature closer to freezing, warming has larger effects on ΔSWE so TP dominates the intensified ablation at low to mid-elevation (500–2000 m). At high elevations with temperature well below freezing, the warmer T in fSST has a negligible impact on ΔSWE. In these regions, precipitation becomes the single dominant factor in snowpack change.
Contributions of each meteorological factor to the winter total ΔSWE. (a) The contribution of each meteorological factor as a function of elevation. The solid lines are the 50% percentile values, and the shadings show the 25%–75% and 5%–95% percentile ranges. (b)–(e) The spatial distribution of these contributions. The values are the relative change of winter ΔSWE (in percentage) when the noted factor is adjusted to fSST scenario. The meanings of these factors are explained in Table 1.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
Figures 8b–e illustrate the spatial distributions of the impact of each factor. The contribution of precipitation is comparable between the north and south, while the contribution of temperature is only notable in the north because of the larger warming and the larger sensitivity of snow ablation to temperature at lower elevation. The difference in dry temperature only has a minor contribution, while the fraction of wet days has no obvious contributions to ΔSWE between the two experiments. The former suggests that temperature affects snowpack in the north mostly through its correlation with precipitation. Warmer TP increases the likelihood of precipitation falling as rain rather than snow, hence contributing to reductions in SWE. Furthermore, under AR conditions, the correlation of warmer temperature and heavier precipitation has a particularly important effect on snow accumulation and melt, with TP having a greater impact than Ptot, likely due to the rain-on-snow effect triggered by warm precipitation. The contributions of these factors to the interannual variability of winter ΔSWE (Fig. S9) have similar patterns as the mean changes (Fig. 8).
The NNm model is trained to predict monthly SWE in winter, and it is validated in Fig. 9 and S10. The model captures 97% of the spatiotemporal variations in winter monthly SWE in the training subset, and over 95% of the variations in the validation subset. Driven by the meteorological conditions of fSST and HIST, NNm correctly predicts the north–south contrast in the mean and variability of monthly SWE as captured by the physics-based WRF simulations. With this model, the impacts of the six meteorological factors are examined and illustrated in Fig. 10. Unlike the winter ΔSWE, all six factors used in NNm have a comparable influence on the monthly snowpack variability. Precipitation increase in fSST leads to a consistent increase of monthly SWE variability in both regions. At the same time, the change in winter T induces a dipole pattern of reduced variability in the north but increased variability in the south. Here we note that the decomposition of the SWE response predicted by the NNm model to the contributions from different meteorological factors does not work as well compared to the NNs model. For example, the sum of the contributions from different meteorological factors to the monthly SWE variation changes in southern Sierra Nevada (Fig. 10) is clearly larger than that predicted by the NNm model (Fig. 9f). This indicates the larger nonlinearity in the relationships between the SWE response and the meteorological factors and the stronger dependence among the meteorological factors on monthly than seasonal time scale.
Validation of NNm model on the estimation of monthly SWE across the western United States. This figure is similar to Fig. 7 but is generated using the monthly SWE data rather than winter total ΔSWE.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
Decomposition of the changes in the variation of monthly SWE. The values are the relative change of monthly SWE variability (in percentage) when the noted factor is adjusted to fSST scenario. The meanings of these factors are explained in Table 2.
Citation: Journal of Hydrometeorology 22, 4; 10.1175/JHM-D-20-0127.1
4. Discussions
a. Implications for mountain hydroclimate in the West Coast
Although the fSST simulation was inadvertently set up with a time-invariant SST field, it provides a useful simulation to understand the impact of local SST warming. More importantly, the relationship between several vital components of the mountain hydroclimate system, i.e., precipitation, temperature, and snowpack can be investigated by comparing fSST with the HIST simulation where time-varying SST was prescribed.
Previous studies have revealed that SST variability associated with large-scale modes of variability accounts for about 20% of the variability of extreme and nonextreme winter precipitation in the U.S. West Coast (Dong et al. 2018). In this study, by controlling the lateral boundary conditions that are influenced by remote SST forcing while changing the local SST in a regional climate simulation, we isolated the impact of local SST forcing on the hydroclimate of the western United States. As shown in Figs. 3d and 3g, warming of the local SST increases the temporal variability of winter precipitation by 10%–20%, despite a complete removal of the local SST temporal variability. Local SST warming increases evaporation from the ocean surface and increases atmospheric moisture. As the lateral boundary conditions in fSST and HIST are identical, an increase in moisture results in an increase in moisture convergence during storm events, thus increasing the precipitation variability in fSST relative to HIST. In the fSST simulation, winter local SST is 2°–3°C warmer than HIST. Based on the same simulations, we demonstrated that precipitation responds to the local SST perturbations at 3–10% K−1 of local SST warming, depending on whether or not precipitation is induced by ARs (Chen and Leung 2020).
Winter extreme precipitation along the West Coast is strongly related to ARs, which significantly modulate the precipitation, temperature, and snowpack (Chen et al. 2019b; Guan et al. 2016). ARs are known to contribute about 40% of annual precipitation along the West Coast (Dettinger et al. 2011). With warming of the local SST, the IVT of ARs over the mountains increases by 5%. This suggests that local SST provides a nontrivial source of moisture for ARs, which may be attributed to the strong winds associated with ARs that increase evaporation from the ocean surface. As revealed by Chen and Leung (2020), the overall atmospheric moisture is increased by 1.2% and 2.5%, respectively, per degree of local SST warming under AR and non-AR conditions. The smaller percentage increase in atmospheric moisture during ARs can be explained by the smaller percentage contribution of local evaporation from the ocean as more moisture is transported from the western boundary of the domain during AR days. Hence, each degree of local SST warming leads to a 3% increase in AR-induced precipitation but a 10% increase in non-AR induced precipitation. The findings of Chen and Leung (2020) and those reported in this study are useful for understanding the impacts of local versus remote oceans under future warming for interpreting analysis of impacts of warming on ARs derived from global climate simulations (Espinoza et al. 2018).
Dacre et al. (2015) found that the water vapor in ARs lost to precipitation is continuously replenished by moisture convergence within the cyclone, suggesting that evaporation from the ocean surface can provide a source of moisture supply for ARs. By modulating the IVT and temperature of ARs, local SST has important effects on AR-induced extreme precipitation and ablation/accumulation of mountain snowpack. Although we focused only on the AR IVT change for the same AR events in the two regional climate simulations, increased atmospheric moisture from the warmer local SST may also increase AR frequency and consequently affect the attribution of the snowpack response to ARs (Chen and Leung 2020). Hence the AR impacts on snowpack presented in our analysis represent likely a lower bound and deserve further investigation in the future using simulations in which the impacts of local SST on the cyclone development of AR storms are better represented than the present model configuration with identical lateral boundary conditions regardless of the local SST.
A key input to NNs and NNm is TP, which reflects the interactions between precipitation and temperature. While previous studies have explored the relationship between temperature, precipitation, and snowpack response, through the common use of Tp, they can be summarized and compared more consistently (Cayan 1996; Li et al. 2020; Ban et al. 2020; Scalzitti et al. 2016; Persson et al. 2011; Barsugli et al. 2020; Wu et al. 2019). We have explored the use of other input variables such as T during precipitation days rather than TP for the NN models, but these variants of NN cannot explain more than 90% of the variations (not shown here) explained by the NNs selected in this study (Fig. 7). Therefore, an accurate description of this interaction is critical to predicting the snowpack response along the West Coast mountains. From this perspective, TP is a useful metric for evaluating meteorological datasets and simulations used in land surface modeling. Meanwhile, TP quantifies how “cold” the precipitation is on average, and Fig. 4 suggests it can serve as a metric of snowpack resistance to warming: if TP is well below freezing, warming may increase snow accumulation. On the contrary, if TP is near or above freezing, warming is likely to reduce snow accumulation. This is consistent with findings from the previous studies on snowpack vulnerability to warming (Qin et al. 2020; Ma et al. 2011; Li et al. 2017; Barsugli et al. 2020; Diffenbaugh et al. 2013; Barnett et al. 2005). Figure 8d further supports the important role of TP in the snowpack response to warming, as it is the only factor that generates a significant north–south difference of winter ΔSWE response, highlighting the contrasting TP in the lower-elevation Cascades range and the higher-elevation Sierra Nevada.
b. Using neural networks to predict snowpack activities
Neural networks have been demonstrated to be a powerful tool to study hydrological processes (Shen 2018; Kratzert et al. 2019). Two neural networks have been established and validated against physics-based simulations. Compared with the traditional approach (such as physics-based land surface modeling), neural networks are highly efficient. Despite the low computational cost, neural networks, especially deep networks, are able to preserve the highly nonlinear relationship between inputs and output, thus providing better performance than linear models. Therefore, as a balanced option between physical and statistical models, NNs have the potential of generating high-quality ensembles fairly quickly.
The two NN models used in this study also demonstrated the relationship between physics complexity and model complexity. At seasonal scale, small daily snow accumulation and ablation largely cancel out, so the winter ΔSWE is dominated by larger daily accumulation/ablation events. These events are likely more strongly correlated with temperature or precipitation: large accumulation is related to heavy precipitation at low temperature; large ablation happens either under high temperature (above freezing) on dry days, or rain-on-snow events where TP is above freezing. Therefore, a four-variable network can predict the seasonal total ΔSWE with satisfactory accuracy.
In comparison, predicting snow activities on a monthly scale is a more complex problem. The ΔSWE of a given month is profoundly affected by the SWE condition at the beginning of the month (SWEi), which in turn depends on the snow activities of the previous month. Therefore, we chose not to estimate ΔSWE but SWEi of each month. SWEi reflects the accumulated impact of precipitation and temperature of multiple previous months. Therefore, a more complex model is required for this task. Previous studies have demonstrated the ability of LSTM in constructing hydrological processes such as rainfall–runoff relationship (Kratzert et al. 2018, 2019; Sahoo et al. 2019; Fang and Shen 2020). Therefore, it is a tool that fits our problem. Compared to the NNs that consist of only 42 tunable parameters, NNm includes 925 tunable parameters, suggesting that a much higher complexity is required to capture the snow behavior at monthly scales.
The networks that are trained in this study can be used to quickly generate large ensembles using other forcings, such as statistically downscaled precipitation and temperature data. Driving NNm with these datasets and comparing the simulated SWE statistics against observation provides an expedient but useful evaluation of these datasets for hydrologic applications. It is noteworthy that prior to selecting the neural network approach, other machine learning methods including random forest regressions and supporting vector machines were tested. However, these methods either cannot achieve the same level of performance as the neural network or fail in simulating the snowpack response under the fSST conditions, despite also being trained using the HIST conditions as in the NN methods.
c. Configuration of the fSST simulation
In the fSST simulation, the local SST pattern is the same throughout the 13-yr simulation period. This setup enables separation of the impacts of local SST from the impacts of remote SSTs, as they are teleconnected as discussed in the introduction, and thus hard to separate in a single climate simulation. Meanwhile, it is necessary to note that the fSST simulation alone does not capture the full impact of local SST, since regional modeling precludes quantification of the potential upstream impact induced by local SST perturbations due to the prescribed lateral boundary conditions. Global models can be used in the future to more fully investigate the direct and upstream impact of local SST perturbations on the mountain snowpack. From this perspective, the fSST experiment addresses a very specific question: Given the same large-scale circulation patterns, how do local SST perturbations affect the precipitation and temperature over the coastal mountains, and thus mountain snowpack?
Another caveat of the fSST simulation is that while the cold season mean local SST is warmer relative to HIST, SST variability is removed. Both of these aspects would affect snowpack. To check the influence of changes in the local SST mean and variability on snowpack variability, random forest regression is used to take the changes of SST mean and standard variation, along with the climatology of local meteorological conditions, and estimate the changes in the daily snowpack variability at a given grid and given month. This model explains 94% of the simulated changes in the snowpack variability. Based on this regression model, we can use the change of local SST mean and variability separately to estimate their impact on the snowpack variability. The results are illustrated in Fig. S11, which shows that the changes of local SST mean increases the snowpack variability across all the elevation bands. This is likely due to a shift in the distribution of land temperature across the freezing point under mean local SST warming, thus increasing the number of melting days and resulting in higher overall snowpack variability. On the contrary, removal of local SST variability only leads to a slight reduction of the snowpack variability. Overall, the simulated snowpack variability change is dominated by the perturbation of local SST mean.
d. Cause of the dipole snowpack response
As indicated by Fig. 2, in the fSST experiment, snowpack in the northern part shows a negative response to warming while the southern part shows a positive response. Since northern local SST experiences a higher increase of local SST (Fig. S1), it is possible that the negative response is due to 1) the amplified warming in the northern mountains associated with the larger SST warming in the northern part of the local ocean or 2) the snowpack system in the north is inherently more sensitive to warming. To check these factors, we computed the sensitivity of air temperature over land to the corresponding local SST warming in the two regions, i.e., the air temperature change over the northern (southern) land per 1 K of local SST warming in the norther (southern) part of the ocean. These sensitivities are illustrated in Fig. S12, and there is no statistically significant difference between the north and south. This indicates that the air temperature response to local SST perturbation in the north is larger than the south because of the larger local SST warming in the north than the south. Therefore, this dipole pattern is caused by 1) air temperature change over land is larger in the north than the south because the local SST in the north warms more than the south and 2) at lower elevation, the north is more sensitive to warming that pushes the temperature across the freezing point.
5. Conclusions
Compared to global or regional SST, the role of nearshore, or local SST in land hydroclimate has been less investigated. In this study, we employed two physics-based regional climate simulations that differ only in the local SST to investigate its role in winter snowpack in the U.S. West Coast mountains. More specifically we compared two regional simulations driven by the same lateral boundary conditions but in one simulation (HIST), SST is prescribed based on observations while in the other simulation (fSST), SST is warmer but has no temporal variability. Based on the simulations, two statistical models (neural networks) were trained and validated to understand and quantify the contributions of different meteorological factors to the winter snowpack response to local SST forcing. Our findings are summarized as follows:
Local SST warming leads to a dipole pattern of snowpack change in winter: a reduction of snowpack in the north (Washington and Oregon) and an increase in snowpack in the south (California).
Increased mean local SST leads to comparable increases in winter precipitation in the Cascades and Sierra Nevada but larger warming of winter mean temperature in the north than the south, with snow–albedo feedback possibly playing a role in the north–south warming contrast. Despite the complete removal of local SST variability, precipitation variability increases because the increased atmospheric moisture with warmer local SST increases moisture convergence during storm events.
The contrasting snowpack response between the north and south is a result of competition between the influence by precipitation and temperature: in the north with mountains at lower elevation, increased temperature increases snowmelt on dry days, increases the likelihood of precipitation falling as rain rather than snow, and triggers more rain-on-snow events, with the latter two contributing more importantly to increasing snow ablation; in the south with mountains at higher elevation, increased precipitation that is more likely in the form of snowfall results in higher snow accumulation.
Though atmospheric rivers are less sensitive to local SST compared to other precipitation events (Chen and Leung 2020), local SST still contributes a nontrivial source of moisture for atmospheric rivers, which exaggerate snow ablation in lower-elevation region, but increases snow accumulation in higher-elevation region. Therefore, atmospheric rivers amplify the dipole responses between the north and south regions.
Neural networks can capture the relationship between precipitation, temperature, and snowpack variations at seasonal and monthly scales. Sensitivity experiments performed with these models indicate that the amount and the average temperature of precipitation are the most critical factors that affect the snowpack response in the mountains of the U.S. West Coast.
The findings highlight the important role of local SST in the mountain hydroclimate of the western United States. They support previous studies on the impact of local SST bias on precipitation bias in the western United States. We further showed that such impacts can extend to biases in the snowpack simulation, with implications for hydrology and water management. Meanwhile, the two neural network structures established in this study (Tables 1 and 2) can inform modeling of snowpack in other snow-covered mountainous regions (e.g., feature selection) or projecting snowpack changes in the U.S. West Coast mountains under climate change.
Acknowledgments
This research is supported by the U.S. Department of Energy Office of Science Biological and Environmental Research as part of the Regional and Global Modeling and Analysis and Multi-Sector Dynamics program areas. The WRF simulation was performed using computing resources of the Pacific Northwest National Laboratory (PNNL) Institutional Computing (PIC) and the National Energy Research Supercomputing Center (NERSC), which is supported by the DOE Office of Science under contract DE-AC02-05CH11231. PNNL is operated for the Department of Energy by Battelle Memorial Institute under contract DE-AC05-76RL01830.
Data availability statement
The regional climate model output used in this study can be found at Zenodo (doi:10.5281/zenodo.4645972). The AR data used in this work can be found at the ARTMIP website (http://www.cgd.ucar.edu/projects/artmip/). The raw WRF output is too large for the public archives and is available upon request to the corresponding authors.
REFERENCES
Alexander, M. A., I. Bladé, M. Newman, J. R. Lanzante, N.-C. Lau, and J. D. Scott, 2002: The atmospheric bridge: The influence of ENSO teleconnections on air–sea interaction over the global oceans. J. Climate, 15, 2205–2231, https://doi.org/10.1175/1520-0442(2002)015<2205:TABTIO>2.0.CO;2.
Ban, C., Z. Xu, D. Zuo, X. Liu, R. Zhang, and J. Wang, 2020: Vertical influence of temperature and precipitation on snow cover variability in the Yarlung Zangbo River basin, China. Int. J. Climatol., 41, 1148–1161, https://doi.org/10.1002/JOC.6776.
Barnett, T. P., J. C. Adam, and D. P. Lettenmaier, 2005: Potential impacts of a warming climate on water availability in snow-dominated regions. Nature, 438, 303–309, https://doi.org/10.1038/nature04141.
Barnston, A. G., and R. E. Livezey, 1987: Classification, seasonality, and persistence of low-frequency atmospheric circulation patterns. Mon. Wea. Rev., 115, 1083–1126, https://doi.org/10.1175/1520-0493(1987)115<1083:CSAPOL>2.0.CO;2.
Barsugli, J. J., A. J. Ray, B. Livneh, C. F. Dewes, A. Heldmyer, I. Rangwala, J. M. Guinotte, and S. Torbit, 2020: Projections of mountain snowpack loss for wolverine denning elevations in the rocky mountains. Earth’s Future, 8, e2020EF001537, https://doi.org/10.1029/2020EF00153.
Caldwell, P., H. N. S. Chin, D. C. Bader, and G. Bala, 2009: Evaluation of a WRF dynamical downscaling simulation over California. Climatic Change, 95, 499–521, https://doi.org/10.1007/s10584-009-9583-5.
Capotondi, A., P. D. Sardeshmukh, E. Di Lorenzo, A. C. Subramanian, and A. J. Miller, 2019: Predictability of US west coast ocean temperatures is not solely due to ENSO. Sci. Rep., 9, 10993, https://doi.org/10.1038/s41598-019-47400-4.
Cayan, D. R., 1996: Interannual climate variability and snowpack in the western United States. J. Climate, 9, 928–948, https://doi.org/10.1175/1520-0442(1996)009<0928:ICVASI>2.0.CO;2.
Chen, X., and L. R. Leung, 2020: Response of landfalling atmospheric rivers on the U.S. West Coast to local sea surface temperature perturbations. Geophys. Res. Lett., 47, e2020GL089254, https://doi.org/10.1029/2020GL089254.
Chen, X., L. R. Leung, Y. Gao, Y. Liu, M. Wigmosta, and M. Richmond, 2018: Predictability of extreme precipitation in western U.S. watersheds based on atmospheric river occurrence, intensity, and duration. Geophys. Res. Lett., 45, 11 693–11 701, https://doi.org/10.1029/2018GL079831.
Chen, X., Z. Duan, L. R. Leung, and M. Wigmosta, 2019a: A framework to delineate precipitation-runoff regimes: Precipitation versus snowpack in the western United States. Geophys. Res. Lett., 46, 13 044–13 053, https://doi.org/10.1029/2019GL085184.
Chen, X., L. R. Leung, M. Wigmosta, and M. Richmond, 2019b: Impact of atmospheric rivers on surface hydrological processes in western U.S. watersheds. J. Geophys. Res., 124, 8896–8916, https://doi.org/10.1029/2019JD030468.
Clark, M. P., M. C. Serreze, and G. J. McCabe, 2001: Historical effects of El Nino and La Nina events on the seasonal evolution of the montane snowpack in the Columbia and Colorado River basins. Water Resour. Res., 37, 741–757, https://doi.org/10.1029/2000WR900305.
Dacre, H. F., P. A. Clark, O. Martinez-Alvarado, M. A. Stringer, and D. A. Lavers, 2015: How do atmospheric rivers form? Bull. Amer. Meteor. Soc., 96, 1243–1255, https://doi.org/10.1175/BAMS-D-14-00031.1.
Dettinger, M. D., D. R. Cayan, H. F. Diaz, and D. M. Meko, 1998: North-South precipitation patterns in western North America on interannual-to-decadal timescales. J. Climate, 11, 3095–3111, https://doi.org/10.1175/1520-0442(1998)011<3095:NSPPIW>2.0.CO;2.
Dettinger, M. D., F. M. Ralph, T. Das, P. J. Neiman, and D. R. Cayan, 2011: Atmospheric rivers, floods and the water resources of California. Water, 3, 445–478, https://doi.org/10.3390/w3020445.
Diffenbaugh, N. S., M. Scherer, and M. Ashfaq, 2013: Response of snow-dependent hydrologic extremes to continued global warming. Nat. Climate Change, 3, 379–384, https://doi.org/10.1038/nclimate1732.
Dong, L., L. R. Leung, F. Song, and J. Lu, 2018: Roles of SST versus internal atmospheric variability in winter extreme precipitation variability along the U.S. West Coast. J. Climate, 31, 8039–8058, https://doi.org/10.1175/JCLI-D-18-0062.1.
Espinoza, V., D. E. Waliser, B. Guan, D. A. Lavers, and F. M. Ralph, 2018: Global analysis of climate change projection effects on atmospheric rivers. Geophys. Res. Lett., 45, 4299–4308, https://doi.org/10.1029/2017GL076968.
Fang, K., and C. Shen, 2020: Near-real-time forecast of satellite-based soil moisture using long short-term memory with an adaptive data integration Kernel. J. Hydrometeor., 21, 399–413, https://doi.org/10.1175/JHM-D-19-0169.1.
Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 5419–5454, https://doi.org/10.1175/JCLI-D-16-0758.1.
Gershunov, A., and D. R. Cayan, 2003: Heavy daily precipitation frequency over the contiguous United States: Sources of climatic variability and seasonal predictability. J. Climate, 16, 2752–2765, https://doi.org/10.1175/1520-0442(2003)016<2752:HDPFOT>2.0.CO;2.
Gershunov, A., T. Shulgina, F. M. Ralph, D. A. Lavers, and J. J. Rutz, 2017: Assessing the climate-scale variability of atmospheric rivers affecting western North America. Geophys. Res. Lett., 44, 7900–7908, https://doi.org/10.1002/2017GL074175.
Goldenson, N., L. R. Leung, C. M. Bitz, and E. Blanchard-Wrigglesworth, 2018: Influence of atmospheric rivers on mountain snowpack in the western United States. J. Climate, 31, 9921–9940, https://doi.org/10.1175/JCLI-D-18-0268.1.
Guan, B., N. P. Molotch, D. E. Waliser, E. J. Fetzer, and P. J. Neiman, 2010: Extreme snowfall events linked to atmospheric rivers and surface air temperature via satellite measurements. Geophys. Res. Lett., 27, L20401, https://doi.org/10.1029/2010GL044696.
Guan, B., D. E. Waliser, R. F. Martin, E. J. Fetzer, and P. J. Neiman, 2016: Hydrometeorological characteristics of rain-on-snow events associated with atmospheric rivers. Geophys. Res. Lett., 43, 2964–2973, https://doi.org/10.1002/2016GL067978.
Gutmann, E. D., R. M. Rasmussen, C. Liu, K. Ikeda, D. J. Gochis, M. P. Clark, J. Dudhia, and G. Thompson, 2012: A comparison of statistical and dynamical downscaling of winter precipitation over complex terrain. J. Climate, 25, 262–281, https://doi.org/10.1175/2011JCLI4109.1.
Hu, J. M., and A. W. Nolin, 2020: Widespread warming trends in storm temperatures and snowpack fate across the Western United States. Environ. Res. Lett., 15, 034059, https://doi.org/10.1088/1748-9326/ab763f.
Hughes, M., J. D. Lundquist, and B. Henn, 2020: Dynamical downscaling improves upon gridded precipitation products in the Sierra Nevada, California. Climate Dyn., 55, 111–129, https://doi.org/10.1007/s00382-017-3631-z.
Jin, J., N. L. Miller, S. Sorooshian, and X. Gao, 2006: Relationship between atmospheric circulation and snowpack in the western USA. Hydrol. Processes, 20, 753–767, https://doi.org/10.1002/hyp.6126.
Kratzert, F., D. Klotz, C. Brenner, K. Schulz, and M. Herrnegger, 2018: Rainfall–runoff modelling using long short-term memory (LSTM) networks. Hydrol. Earth Syst. Sci., 22, 6005–6022, https://doi.org/10.5194/hess-22-6005-2018.
Kratzert, F., M. Herrnegger, D. Klotz, S. Hochreiter, and G. Klambauer, 2019: NeuralHydrology – Interpreting LSTMs in hydrology. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, W. Samek et al., Eds., Springer, 347–362.
Li, D., M. L. Wrzesien, M. Durand, J. Adam, and D. P. Lettenmaier, 2017: How much runoff originates as snow in the western United States, and how will that change in the future? Geophys. Res. Lett., 44, 6163–6172, https://doi.org/10.1002/2017GL073551.
Li, Z., Y. Chen, Y. Li, and Y. Wang, 2020: Declining snowfall fraction in the alpine regions, Central Asia. Sci. Rep., 10, 3476, https://doi.org/10.1038/s41598-020-60303-z.
Linkin, M. E., and S. Nigam, 2008: The North Pacific Oscillation–West Pacific teleconnection pattern: Mature-phase structure and winter impacts. J. Climate, 21, 1979–1997, https://doi.org/10.1175/2007JCLI2048.1.
Ma, L., D. Qin, L. Bian, C. Xiao, and Y. Luo, 2011: Assessment of snow cover vulnerability over the Qinghai-Tibetan Plateau. Adv. Climate Change Res., 2, 93–100, https://doi.org/10.3724/SP.J.1248.2011.00093.
McCabe, G. J., M. P. Clark, and L. E. Hay, 2007: Rain-on-snow events in the western United States. Bull. Amer. Meteor. Soc., 88, 319–328, https://doi.org/10.1175/BAMS-88-3-319.
McPhaden, M. J., and X. Yu, 1999: Equatorial waves and the 1997–98 El Niño. Geophys. Res. Lett., 26, 2961–2964, https://doi.org/10.1029/1999GL004901.
Mejia, J. F., D. Koračin, and E. M. Wilcox, 2018: Effect of coupled global climate models sea surface temperature biases on simulated climate of the western United States. Int. J. Climatol., 38, 5386–5404, https://doi.org/10.1002/joc.5817.
Mesinger, F., and Coauthors, 2006: North American Regional Reanalysis. Bull. Amer. Meteor. Soc., 87, 343–360, https://doi.org/10.1175/BAMS-87-3-343.
Neiman, P. J., L. J. Schick, F. M. Ralph, M. Hughes, and G. A. Wick, 2011: Flooding in western Washington: The connection to atmospheric rivers. J. Hydrometeor., 12, 1337–1358, https://doi.org/10.1175/2011JHM1358.1.
Newell, R. E., N. E. Newell, Y. Zhu, and C. Scott, 1992: Tropospheric rivers? – A pilot study. Geophys. Res. Lett., 19, 2401–2404, https://doi.org/10.1029/92GL02916.
Persson, A., P. L. Langen, P. Ditlevsen, and B. M. Vinther, 2011: The influence of precipitation weighting on interannual variability of stable water isotopes in Greenland. J. Geophys. Res., 116, D20120, https://doi.org/10.1029/2010JD015517.
Persson, P. O. G., P. J. Neiman, B. Walter, J.-W. Bao, and F. M. Ralph, 2005: Contributions from California coastal-zone surface fluxes to heavy coastal precipitation: A CALJET case study during the strong El Niño of 1998. Mon. Wea. Rev., 133, 1175–1198, https://doi.org/10.1175/MWR2910.1.
Potito, A. P., D. F. Porinchu, G. M. MacDonald, and K. A. Moser, 2006: A late Quaternary chironomid-inferred temperature record from the Sierra Nevada, California, with connections to northeast Pacific sea surface temperatures. Quat. Res., 66, 356–363, https://doi.org/10.1016/j.yqres.2006.05.005.
Qin, Y., and Coauthors, 2020: Agricultural risks from changing snowmelt. Nat. Climate Change, 10, 459–465, https://doi.org/10.1038/s41558-020-0746-8.
Ralph, F. M., and M. D. Dettinger, 2011: Storms, floods, and the science of atmospheric rivers. Eos, Trans. Amer. Geophys. Union, 92, 265–266, https://doi.org/10.1029/2011EO320001.
Ralph, F. M., P. J. Neiman, G. A. Wick, S. I. Gutman, M. D. Dettinger, D. R. Cayan, and A. B. White, 2006: Flooding on California’s Russian River: Role of atmospheric rivers. Geophys. Res. Lett., 33, L13801, https://doi.org/10.1029/2006GL026689.
Ralph, F. M., J. J. Rutz, J. M. Cordeira, M. Dettinger, M. Anderson, D. Reynolds, L. J. Schick, and C. Smallcomb, 2019: A scale to characterize the strength and impacts of atmospheric rivers. Bull. Amer. Meteor. Soc., 100, 269–289, https://doi.org/10.1175/BAMS-D-18-0023.1.
Rasmussen, R., and Coauthors, 2011: High-resolution coupled climate runoff simulations of seasonal snowfall over Colorado: A process study of current and warmer climate. J. Climate, 24, 3015–3048, https://doi.org/10.1175/2010JCLI3985.1.
Redmond, K. T., and R. W. Koch, 1991: Surface climate and streamflow variability in the western United States and their relationship to large-scale circulation indices. Water Resour. Res., 27, 2381–2399, https://doi.org/10.1029/91WR00690.
Sagarika, S., A. Kalra, and S. Ahmad, 2016: Pacific Ocean SST and Z500 climate variability and western U.S. seasonal streamflow. Int. J. Climatol., 36, 1515–1533, https://doi.org/10.1002/joc.4442.
Sahoo, B. B., R. Jha, A. Singh, and D. Kumar, 2019: Long short-term memory (LSTM) recurrent neural network for low-flow hydrological time series forecasting. Acta Geophys., 67, 1471–1481, https://doi.org/10.1007/s11600-019-00330-1.
Scalzitti, J., C. Strong, and A. Kochanski, 2016: Climate change impact on the roles of temperature and precipitation in western U.S. snowpack variability. Geophys. Res. Lett., 43, 5361–5369, https://doi.org/10.1002/2016GL068798.
Shen, C., 2018: A transdisciplinary review of deep learning research and its relevance for water resources scientists. Water Resour. Res., 54, 8558–8593, https://doi.org/10.1029/2018WR022643.
Shields, C. A., and Coauthors, 2018: Atmospheric River Tracking Method Intercomparison Project (ARTMIP): Project goals and experimental design. Geosci. Model Dev., 11, 2455–2474, https://doi.org/10.5194/gmd-11-2455-2018.
Waliser, D., and B. Guan, 2017: Extreme winds and precipitation during landfall of atmospheric rivers. Nat. Geosci., 10, 179–183, https://doi.org/10.1038/ngeo2894.
Wise, E. K., 2010: Spatiotemporal variability of the precipitation dipole transition zone in the western United States. Geophys. Res. Lett., 37, L07706, https://doi.org/10.1029/2009GL042193.
Wrzesien, M. L., M. T. Durand, T. M. Pavelsky, I. M. Howat, S. A. Margulis, and L. S. Huning, 2017: Comparison of methods to estimate snow water equivalent at the mountain range scale: A case study of the California Sierra Nevada. J. Hydrometeor., 18, 1101–1119, https://doi.org/10.1175/JHM-D-16-0246.1.
Wu, S., X. Zhang, J. Du, X. Zhou, Y. Tuo, R. Li, and Z. Duan, 2019: The vertical influence of temperature and precipitation on snow cover variability in the central Tianshan Mountains, northwest China. Hydrol. Processes, 33, 1686–1697, https://doi.org/10.1002/hyp.13431.