An analysis of annual hydroclimatic variability in the Upper Colorado River basin (UCRB) for the period of 1906–2006 was performed to understand the dominant modes of multidecadal variability. First, wavelet-based spectral analysis was employed for streamflow at Lees Ferry, Arizona (aggregate location for UCRB flow), which identified two significant modes: a “low frequency” (~64-yr period) mode and a strong “decadal” (~15-yr period) component active only in recent decades. Subsequent investigation of temperature and precipitation data for the UCRB indicated that the low-frequency variability is associated with temperature via modulation of runoff efficiency while the decadal is strongly tied to moisture delivery. Simple hydrology and climate model experiments are also provided to support the aforementioned findings.
Correlation of UCRB precipitation with global sea surface temperature (SST) anomalies showed a strong link with the equatorial and northern Pacific during periods of heightened variability of the decadal mode. The correlation of UCRB temperature with global SST anomalies showed strongest values in the Atlantic consistent with the Atlantic multidecadal oscillation mode. Wavelet spectral analysis of paleo-reconstructed streamflow at Lees Ferry shows both the low-frequency and decadal flow variability features. Furthermore, the strength of the decadal mode is modulated at an ~75-yr time scale, and these are consistent with epochal variations of overall streamflow variance.
In many regions of the United States, water management and reservoir operations occur at the subannual time scale. These projects often have seasonal targets to accomplish objectives such as flood control, recreation, and hydropower. Thus, for facilities with limited storage capacity, information beyond a seasonal forecast is of little use. On the other hand, the Colorado River basin boasts a collective storage capacity of approximately four times the mean annual flow of the river. Furthermore, over 80% of that storage is found in Lakes Powell and Mead alone. Thus, understanding persistent modes of streamflow variability at the interannual and longer time scale is of great importance for water management in this system.
There is increasing evidence that the western U.S. hydroclimate exhibits significant interannual (year-to-year) variations, driven by large-scale climate features such as the ENSO, Pacific decadal oscillation (PDO), and Atlantic multidecadal oscillation (AMO). Particularly, the links to ENSO are robust, widespread, and have a long history of recognition. During El Niño events (warm sea surface temperature anomalies in the central and eastern equatorial Pacific Ocean), the winter subtropical jet over the southwestern United States strengthens (Horel and Wallace 1981) and consequently the Pacific Northwest experiences below-normal and the Southwest above-normal precipitation (Cayan et al. 1998; Dettinger et al. 1998; Diaz and Markgraf 1992, 2000; Redmond and Koch 1991). Generally the opposite effect is observed during La Niña events (cooler sea surface temperature in the central and eastern equatorial Pacific), but some nonlinearities are present in this teleconnection (Clark et al. 2001; Hoerling et al. 1997; Rajagopalan et al. 2000). Similar ENSO teleconnection patterns have been observed in the interannual variability of winter snow water equivalent (Cayan 1996; Clark et al. 2001), surface temperature (Gershunov and Barnett 1998; Higgins et al. 2002; Redmond and Koch 1991), and streamflow (Dracup and Kahya 1994; Grantz et al. 2005; Hamlet and Lettenmaier 1999; Kahya and Dracup 1994; Maurer et al. 2004; Piechota et al. 1997; Regonda et al. 2006; Tootle et al. 2005). Shifts in streamflow seasonality (Cayan et al. 2001; Grantz et al. 2007; Regonda et al. 2005; Stewart et al. 2005) have also been linked to ENSO and a general warming trend in recent decades. At decadal time scales, the PDO is often considered the primary driver of hydrologic variability (Brown and Comrie 2004; Hidalgo and Dracup 2003; McCabe and Dettinger 1999, 2002), despite debate on the independence of the PDO from ENSO (Newman et al. 2003).
Hunter et al. (2006) examined the link between 1 April snow water equivalent at Snotel (http://www.wcc.nrcs.usda.gov/snow/about.html) sites across the western United States and various climate indices. They found that for regions in the Upper Colorado River basin (UCRB), La Niña years and the negative phase of AMO corresponded with above-average snowpack. Additionally, the cold phase of PDO, when compared with the warm, was found to correspond with reduced snowpack in similar regions. Timilsena et al. (2009) noted similar climatic ties to streamflow in the UCRB, finding the warm phase of PDO to be associated with above-average flow. Also, a weak inverse relationship between flow and the AMO phase was noted. However, El Niño years were linked with above-average flow for much of the basin, albeit more pronounced in the lower basin. Timilsena et al. (2009) also found that the effects of ENSO and PDO were intensified when in phase and suppressed when out of phase.
Thomas (2007) studied Lower Colorado River basin (LCRB) flow ties to the same indices at seasonal and annual scales, for a variety of lead/lag times. The ENSO and PDO were both found to have strong relationships for winter/spring and annual flow volumes at little to no lag. An inverse relationship was detected for annual and winter/spring volumes with the AMO at a lag of 6–24 months. McCabe et al. (2007) examined decade-scale variability in global sea surface temperatures (SSTs) and UCRB flow and also found strong ties to the AMO and PDO. Furthermore, from analysis of climate division temperature and precipitation data, for much of the United States, positive (negative) AMO corresponds with above (below) average temperatures and below (above) average precipitation.
While these works offer considerable insight into the influence of various climatic phenomena on the hydroclimate of the intermountain west region, little has been done from the perspective of detection and attribution for significant streamflow variability modes in the Colorado basin. To this end, we performed an analysis of historical and paleo-reconstructed Colorado River streamflow using wavelet spectral methods. Subsequently, physical processes and potential climate phenomena were identified and associated with interannual and longer modes of variability. This approach is similar to that employed by Coulibaly and Burn (2004) for basins across Canada and by Massei et al. (2011) in the Mississippi River drainage. While some findings presented in this study have been previously examined, this work presents a novel and comprehensive analysis of observed and paleo-reconstructed Colorado River flow variability modes, offers an explanation of these modes from hydrologic and climatic perspectives, and finally uses idealized modeling experiments to support and further explore these findings. The paper is organized as follows: datasets and methods used are next presented, followed by results, and discussion.
The following provides an overview of the various datasets used in the subsequent analyses of this chapter.
a. Colorado River natural flow
In this work, the annual water-year (October–September) natural streamflow data at Lees Ferry on the Colorado River for the period 1906–2006 are employed. Naturalized streamflow data are computed by removing anthropogenic impacts (i.e., reservoir regulation, consumptive water use, etc.) from the recorded historical flows (United States Department of the Interior 2005). These data are developed and updated regularly by the Bureau of Reclamation. Figure 1 shows the Colorado River basin. Of significance is the Lees Ferry gauge, which divides the basin into upper and lower portions per the 1922 Colorado River Compact (Colorado River Commission et al. 1923). More than 90% of the natural streamflow in the basin passes through this point. As such, we chose to perform our analysis at this location (United States Department of the Interior 2005).
b. Climate index data
The AMO index is computed as a monthly area-weighted average of North Atlantic (latitudes north of the equator) sea surface temperatures, which is subsequently detrended. Gridded data used for this calculation are from the Kaplan SST dataset. These data are updated monthly and have a spatial resolution of 5° × 5°. Values were accessed from the NOAA Physical Sciences data website (http://www.esrl.noaa.gov/psd/data/timeseries/AMO/) and are available for the period 1856–present.
c. Upper Colorado River basin precipitation and temperature data
Monthly precipitation on a high-resolution spatial grid (4 km × 4 km) is available for the period 1895–present, known as the Precipitation–Elevation Regression on Independent Slopes Model (PRISM) dataset. The PRISM method is a statistical interpolation approach that utilizes point measurements in conjunction with digital elevation data. While originally developed for precipitation data, the method has been successfully applied to other climate variables such as temperature and snowfall (Daly et al. 1994). Data and additional information are available at the Oregon State University Oregon Climate Service website (http://www.prism.oregonstate.edu/.)
d. Global sea surface temperatures
Monthly global sea surface temperature anomalies on a 5° × 5° scale for the period 1856–present were obtained via the Lamont-Doherty Earth Observatory/International Research Institute website data library (http://portal.iri.columbia.edu/portal/server.pt). The data are produced from a “reduced space” optimal smoother algorithm applied to global SST monthly anomalies obtained from the U.K. Hadley Center archives (Bottomley et al. 1990; Kaplan et al. 1998). This dataset is regularly updated and used extensively in climate diagnostics studies.
e. Paleo-reconstructed data
Paleo-reconstructed streamflow and climate data from tree rings were employed in this work. Specifically, water year (October–September) streamflow at Lees Ferry, Arizona (Woodhouse et al. 2006), spanning the period 1490–1997 and a reconstructed Atlantic multidecadal oscillation (Gray et al. 2004) covering the period 1572–1985 were obtained from the NOAA paleoclimatology web page (http://www.ncdc.noaa.gov/paleo/paleo.html). The reconstruction process typically involves fitting a regression model to the tree ring data and the variable of interest (e.g., streamflow, AMO index, etc.) during the modern period when both datasets are available. The fitted regression model is then applied to tree ring data of the preobservation period to obtain reconstructed estimates of the desired variables. The streamflow reconstruction was derived from tree ring chronologies in the Colorado basin, while the AMO reconstruction was developed with records from eastern North America, Western Europe, Scandinavia, and the Middle East. For detailed discussion of the reconstruction methods, the readers are referred to the aforementioned references and others therein.
The objective of this work was to obtain further insight into the source and characteristics of UCRB flow variability; thus, spectral analyses were first performed. Specifically, wavelet analyses of streamflow, temperature, and precipitation data were used to identify dominant modes of variability in the hydroclimate of the Upper Colorado River basin. To highlight and isolate the impact of various factors on streamflow, a combination of wavelet bandpass filtering and simple linear regression were employed. Subsequently, the component time series were correlated with global sea surface temperature data to aid in identifying the drivers of the variability. Last, paleo-reconstructed data were analyzed to investigate persistence of climatic ties and long-term fluctuations of the dominant periodicities.
Wavelet spectral analysis
The use of wavelet spectral analysis to identify dominant spectral peaks and their temporal variations in geophysical data is increasingly popular (Addison 2002; Kwon et al. 2009; Torrence and Compo 1998; Torrence and Webster 1998). Wavelet methods have several advantages over traditional spectral estimation techniques, such as their ability to effectively capture “local” features resulting from amplitude modulation. Additionally, they are efficient at revealing quasiperiodic signals in data in the presence of considerable noise, which is often the case with geophysical data. In this work, several wavelet-based methods were employed, a brief overview of these methods is provided below.
For the purpose of investigating modes of variability, the local and global wavelet power spectra were analyzed. The local power spectrum shows variability in the two dimensional, frequency–time space. By averaging across the temporal domain at each scale, the global power spectrum is computed. For both local and global spectra a 90% significance white noise background spectra level was employed. Additionally, a type of wavelet bandpass filtering was used to generate scale-specific components of the original time series. In simplest terms, this is a back-transformation of the wavelet-transformed data to the original space over a limited range of variability scales. This work employs the Morlet wavelet (ω0 = 6) with a starting scale of 2δt and a scale width (δj) of 0.25. Additionally, the data have been “padded” with zeros to reduce edge effects. For a detailed discussion of the wavelet methods, the reader is referred to A Practical Guide to Wavelet Analysis (Torrence and Compo 1998).
a. Dominant modes of variability
To begin, the global and local wavelet power spectra were computed for annual water-year flow at Lees Ferry, Arizona (Fig. 2). This location has twofold significance; it serves as the delineation point between the upper and lower basins per the 1922 Colorado River Compact (Colorado River Commission. et al. 1923) and is the hydrologic aggregate for over 90% of the total basin flow. Significant features in both global and local spectra were identified against a white noise background spectrum at the 90% confidence level. From visual inspection of the global spectrum, there are two obvious, significant features—a peak in the 8–16-yr period and a lower frequency (~64 year period) feature. An assessment of the local spectrum indicated that the low-frequency variability is persistent throughout the time domain, while the “decadal” feature is only active in the most recent 30 years. Similar findings were shown in Hidalgo and Dracup (2003) through their use of wavelet analysis for various hydroclimatic data pertinent to the Colorado basin.
To better understand the source of these variability modes, the same wavelet spectral analysis was performed for Upper Colorado River basin temperature and precipitation data. Figure 3 shows the global and local wavelet power spectra for PRISM upper basin precipitation data. The major feature is power in the 8–16-yr period, which is only active in the most recent 3–4 decades and is strikingly similar to the decadal variability seen in the streamflow data.
As a complement, Fig. 4 shows the wavelet power spectra for upper basin PRISM temperature data. These data are characterized by persistent, low-frequency (~64-yr period) variability and bear considerable similarity to the low-frequency component of the streamflow data. Thus, subsequent analyses were based on the hypothesis that low-frequency variability in Colorado River flow is associated with temperature and recent decadal variability with moisture delivery.
While investigating the link between precipitation and temperature with flow, it is important to recognize that observational temperature and precipitation data are not independent. Surface moisture availability is a key factor in the hydrologic energy balance. In simplest terms, evaporation has a cooling effect due to the energy expenditure associated with latent heat of vaporization. In a dry year, the evaporation component of the hydrologic cycle is moisture limited and excess energy results in additional surface warming. Likewise, relative to years with below-average precipitation, wet years tend to be cooler owing to enhanced evaporation. Thus, to remove this interdependency, a linear regression was fit between temperature and precipitation. Residuals from the model were retained as temperature data independent of precipitation. Results indicated that 11% of the temperature variance is explained by precipitation. For all subsequent analyses and results, the temperature data employed have the influence of precipitation removed, so as to avoid confounding the influences of these two variables, and are referred to as “residual temperature.”
To highlight the relationship between flow, residual temperature, and precipitation, time series plots are provided (Figs. 5 and 6). Precipitation and flow have a strong positive relationship (correlation = 0.77), while residual temperature has a slightly weaker and negative association (correlation = −0.32).
Figure 7 offers scatterplots of precipitation versus flow and residual temperature versus flow as additional evidence of the respective relationships. These links are quite significant; in fact, a simple multivariate linear regression using residual temperature and precipitation as dependent variables explains about 70% of the flow variability.
From a mass balance perspective, precipitation represents the maximum potential runoff (assuming 100% basin runoff efficiency), and residual temperature likely contributes to modulation of runoff efficiency. To explore this further, annual UCRB runoff efficiency was computed as the annual flow volume divided by annual precipitation volume. This time series is shown in Fig. 8. From analysis of these data, a weak tie to precipitation was seen (correlation = 0.29), which is likely due to a few “outlier” precipitation years. In these instances, an especially below average precipitation year can cause efficiency to suffer. However, for most years, including particularly wet ones, precipitation magnitude has little to no link with runoff efficiency (Fig. 9).
Residual temperature, on the other hand, has a much stronger and robust relationship with runoff efficiency (correlation = −0.49). As shown in Fig. 10, efficiency steadily declines with increasing residual temperature. Furthermore, the slope from a linear regression fit to data in Fig. 10 indicates that per degree warming, runoff percent efficiency decreases by 2. Thus, assuming a basin-average efficiency of 14.5%, one degree warming should be expected to decrease annual streamflow by 13.8% or approximately 2 MaF. A number of previous studies have investigated the impact of temperature on streamflow and runoff efficiency across the United States, many focusing on the U.S. Southwest and Colorado River basin (Christensen and Lettenmaier 2007; Langbein 1949; Nash and Gleick 1991; Revelle and Waggoner 1983). While the results range somewhat from study to study, the findings presented here are generally consistent with previous work. All of these investigations suggest that warming conditions will result in either reduced streamflow or reduced runoff efficiency. Thus, it should be quite evident that understanding temperature and precipitation variability modes is critical to the overall hydrology in the Upper Colorado River basin.
b. Links to large-scale climate forcings: Low-frequency mode
To help understand the links between UCRB hydroclimatic variability and large-scale climate forcings, the previously identified major drivers of flow (water year temperature and precipitation) were correlated with winter season global SST anomalies (http://iridl.ldeo.columbia.edu/index.html). Prior to the correlation analysis, residual temperature and precipitation data were filtered using wavelet methods to highlight their respective key variability scales. The residual temperature correlation was first performed using a low-pass filter (periods > 32 yr), and results are shown in Fig. 11. This figure shows significant correlation in the North Atlantic and portions of the Pacific Ocean. This pattern is distinctly reminiscent of the Atlantic multidecadal oscillation (Enfield et al. 2001). Emerging research suggests an association between the AMO and western U.S. hydroclimate (Hidalgo 2004; McCabe et al. 2007, 2008), raising the question about hopeful prospects for skillful long-lead projection of streamflow.
To further highlight this correlation, Fig. 12 shows a scaled AMO index plotted with scaled low-pass wavelet filtered UCRB residual temperature data. While the low-frequency nature of the signal offers less than two cycles over the period of record, the tracking of the two datasets is quite striking. McCabe et al. (2008) showed similar covariance between the AMO and temperature variations across much of the continental United States.
c. Links to large-scale climate forcings: Decadal mode
The same SST correlation was performed with the UCRB precipitation data with a high-pass wavelet filter (periods < 20 yr). It is well established that moisture delivery to the western United States is Pacific in origin. From Fig. 13, the majority of regions with significant correlation are found in the Pacific, albeit relatively weak in magnitude and sparse in extent. The most noticeable feature in the map is a relatively large area of positive correlation in the equatorial Pacific, reminiscent of the ENSO pattern. This is not surprising; as previously discussed, there is a wealth of literature linking this phenomenon with the hydroclimate of the western United States.
Given the substantial nonstationary decadal variance seen in both precipitation and flow data, streamflow wavelet power spectra are revisited. Recall that the majority of precipitation-associated variability is seen in recent decades and not persistent throughout the time domain, in direct contrast with the lower frequency temperature variability. Thus, the SST/precipitation correlation was repeated for the period 1970–2006. The resulting correlation map shows heightened Pacific correlation—bearing semblance to ENSO and possibly the Pacific decadal oscillation (Diaz and Markgraf 2000; Mantua and Hare 2002; Mantua et al. 1997; White and Cayan 2000) (Fig. 14). It should be noted that, owing to the shortened length of the subperiod, 90% significant correlation values are greater than 0.26, whereas for the entire period of record this value is 0.16. As a complementary plot, Fig. 15 shows the high-pass filtered temperature data correlated with SST data for the period 1906–69. There are few regions with significant values (0.2 correlation), and they appear to be quite random. Suffice it to say, any link between UCRB precipitation and the equatorial Pacific is almost exclusively from the past ~30 years.
Interestingly, there is strong evidence of heightened activity in the Pacific beginning in the early to mid-1970s (Graham 1994; Hare and Mantua 2000; Miller et al. 1994; Trenberth and Hurrell 1994). This period is characterized by a shift in ENSO characteristic, a strong, positive PDO regime, and a persistent negative North Pacific index (NPI) (An and Wang 2000; Diaz et al. 2001; Latif and Barnett 1994; Rajagopalan et al. 1997; Torrence and Webster 1999; Wang 1995). Furthermore, post-1975, precipitation variance in the UCRB increased by 50% relative to pre-1975 period, and the relationship between flow and precipitation strengthened post-1975. For the period prior to 1975, precipitation explains 55% of the flow variance, while post-1975 it explains over 70%. In light of this, understanding the duration and frequency of periods characterized by heightened Pacific activity, and thus precipitation, is of great interest.
d. Temporal variability over paleo time period
Confident detection and attribution of spectral features, particularly lower frequency (such as the ~64-yr period) can be challenging due to limited data, boundary effects, and fewer degrees of freedom. In the case of the recent period of strong decadal variance, it is difficult to know if this is an anomalous epoch or a feature that waxes and wanes over time. Paleo-reconstructed data covering a longer period provide a good source to further investigate these uncertainties. The Woodhouse et al. (2006) reconstruction of upper basin streamflow was utilized. The wavelet spectrum of this paleostreamflow (Fig. 16) shows similar spectral signals as that of the dominant mode of streamflow (Fig. 2). In particular, two interesting features: (i) the significant power of an ~64-yr period is robust over the entire duration, except during 1700–1800 when it exhibits an ~32-yr period, and (ii) the significant decadal signal (~8–16-yr period) exhibits epochal behavior. Interestingly, in Hidalgo (2004), variability bands of reconstructed Palmer drought severity index (PDSI) data for the southwestern United States were examined and a similar weakening of low-frequency variability strength during the eighteenth century was noted.
Low-frequency reconstructions of the paleostreamflow in the 64–80-yr band were computed and shown with the paleo-AMO index (Gray et al. 2004) in Fig. 17. They covary with an inverse relationship reasonably well over the entire period, with increased amplitudes in the recent centuries and a general weakening during the eighteenth century. This provides further corroboration of the low-frequency covariance of streamflow with the AMO identified earlier in this research and by others (Hunter et al. 2006; McCabe et al. 2008).
To investigate the temporal variability of the decadal band (~8–16-yr period) signal, the scale-averaged wavelet power (SAWP) was computed, which represents the strength of this signal over the time domain (Torrence and Compo 1998). The epochal nature of the variability is quite apparent, and a 25-yr moving window variance of the paleostreamflow (Fig. 18) shows that epochs with enhanced SAWP are consistent with increased overall streamflow variance (correlation = 0.75). The recent decades appear to be one such epoch.
A global spectrum of the SAWP (Fig. 19) shows two distinct peaks, suggesting that the decadal variability signal tends to be modulated at ~75- and ~150-yr period. Perhaps only coincidence, it is worth noting that the decadal variability strength is modulated at approximately the same time scale as the overall low-frequency component (64–80-yr period). It is possible that there may be some interplay between these components, but without additional work this is largely speculative. The ~150-yr peak is more difficult to attribute and may be largely due to the magnitude of the peak in decadal strength around 1850.
e. Additional climate model results
To further investigate the low-frequency relationship between Atlantic sea surface temperatures and the Colorado River basin, results from idealized climate model experiments are presented. Schubert et al. (2009) used five general circulation models (GCMs) to examine the impact of large-scale sea surface temperature patterns on drought in the United States. Specifically, the National Center for Atmospheric Research’s Community Climate Model version 3.0 (CCM3) and Community Atmospheric Model version 3.5 (CAM3.5) were both used in addition to the NASA Seasonal-to-Interannual Prediction Project model version 1 (NSIPP-1). Also, the NOAA Global Forecast System model (GFS) and Atmospheric Model version 2.1 (AM2.1) were employed. For model runs forced with a neutral Pacific and warm Atlantic Ocean pattern, four of the five GCMs showed above-average annual temperatures for the continental United States, with these anomalies ranging from 0.15° to 0.4°C. The Atlantic SST forcing pattern for these experiments (not shown) strongly resembles the positive phase of the Atlantic multidecadal oscillation. Figure 20 shows additional results based on the original work of Schubert et al. (2009), highlighting the impact of a warm versus cool Atlantic Ocean on annual temperature and 200-mb geopotential height anomalies. Warming is most elevated across the Great Plains region, but moderate increases are noted for the Colorado basin headwaters. These results offer some support for the hypothesis that the AMO may have a hand in UCRB low-frequency temperature variance and consequently runoff efficiency. However, additional work is needed to establish the explicit physical mechanism of such a relationship.
f. Additional hydrology model results
Further analysis regarding the link between temperature/precipitation and flow variability modes was conducted using a simple hydrology model developed by the U.S. Geological Survey (Wolock and McCabe 1999a,b). In this, the model was first forced with the same Upper Colorado River basin PRISM temperature and precipitation data described earlier. The model-derived annual streamflow for Lees Ferry was compared with that of the natural flow dataset for the period of 1906–2006 for validation purposes. While the modeled flow showed a wet bias (~3 MaF yr−1), the two time series were highly correlated (>0.95). Additionally, the wavelet power spectra of these time series show similar features. Thus, the model sufficiently captures year to year variability for the purpose of this work.
To isolate the temperature and precipitation influences on streamflow, model runs with the following inputs were made: 1) PRISM temperature data and precipitation climatology (i.e., average annual data from the historic period) and 2) PRISM precipitation data and temperature climatology. The wavelet power spectra were computed for these two scenarios (Fig. 21). From these figures, the modeled flow data from PRISM temperature and precipitation climatology show a marked low-frequency variability mode. The data from the model forced with temperature climatology and PRISM precipitation displayed the same decadal variability seen in the natural flow time series from ~1975 to present. There is a “low-frequency” component in this spectrum; however, it is a longer period than that seen in the historic natural flow data and may be an artifact of a small decreasing trend in the precipitation data. The result of this exercise is through a physically based hydrology model, it was confirmed that low-frequency Colorado River streamflow variability is due to similar low-frequency temperature oscillations and that decadal flow variations are a direct result of precipitation.
An analysis of UCRB hydrologic variability was performed to better understand the dominant modes of streamflow variability. Wavelet spectral analysis of annual Lees Ferry streamflow indicated two dominant variability scales—a persistent low-frequency component (~64-yr period) and a nonstationary decadal signal (8–16-yr period). Subsequent spectral analysis suggested that the low-frequency component of flow is associated more with temperature variability, while the higher frequency is associated more with precipitation variability.
A strong decadal signal was seen in recent decades for both flow and precipitation and was quite weak for the first two-thirds of the historical record. Correlation of high-pass filtered precipitation data with global sea surface temperatures over the entire record (1906–2006) indicated a weak, but significant tie to the equatorial Pacific. The same analysis was performed for the period of high decadal variability (1970–2006) and showed a heightened equatorial Pacific link, while there was little to report for 1906–69. Enhanced decadal variability in precipitation and streamflow coincided with well-documented shifts in Pacific Ocean characteristics; the PDO switched from a period of negative to positive phase, while ENSO had a tendency toward stronger, persistent El Niño events and weaker, less frequent La Niñas.
Temperature oscillations on the 64-yr period scale were shown to modulate runoff efficiency. Per degree warming, a 14% decrease in UCRB discharge was implied by the diagnosis of historical Lees Ferry streamflow and climate data. From correlation of low-pass-filtered temperature data with global SST data, a strong link to the Atlantic was noted. Further investigation showed distinct covariance with the AMO index, also supported by results from climate model experiments. As no causal relationship has been established at this time, further research is needed to assess the reliability of the aforementioned association. Specifically, identification of a physical mechanism by which the Colorado River basin is influenced by Atlantic SSTs will solidify the proposed connection. As anthropogenic climate change is all but certain to warm the UCRB, continued study of this variability mode will be critical for understanding future flow characteristics, in terms of both magnitude and variability.
Paleo-reconstructed streamflow at Lees Ferry was analyzed to gain insight to the persistence and strength of variability features identified over the historical period. Decadal variance appears to wax and wane in strength at roughly a ~75-yr time scale. Furthermore, the epochal variations of the strength of the decadal signal are consistent with epochal variations of overall streamflow variance. When the decadal signal is active, the streamflow exhibits enhanced variance. Low-frequency variance was also seen throughout much of the paleodata, with a lull/shift to higher frequency variability in the 1700s. A comparison of the AMO reconstruction of Gray et al. (2004) and low-pass-filtered paleoflow at Lees Ferry showed reasonable history of the inverse flow/AMO relationship from the observed period.
Additionally, simple modeling experiments support these findings. For coupled GCMs forced with “warm AMO” Atlantic and neutral Pacific SST patterns, warming was seen over parts of the Upper Colorado River basin, consistent with the relationship seen in Fig. 12. The hydrology model exercise demonstrated that the significant spectral features of input temperature and precipitation data translate to similar scales of flow variability.
In closing, many operational and planning decisions for water resources management are at the multidecadal horizon. Thus, robust flow scenarios at these time scales are important for providing accurate estimates of system risk and to devise optimal management strategies. The findings from this research could lead to improved stochastic streamflow projection techniques. Stochastic flow generation methods for near term (i.e., 2–3 decades) do not incorporate spectral characteristics, and consequently, the scenarios tend to provide an inaccurate estimate of system risk.
Funding for this research by Bureau of Reclamation and the Western Water Assessment RISA program at the University of Colorado is gratefully acknowledged. We thank Jon Eischeid and Philip Pegion for their assistance in the modeling portions of this work. Thanks are also due to the Center for Advanced Decision Support in Water and Environmental Systems (CADSWES) at the University of Colorado, Boulder, for use of its facilities and computational support. We also thank the three reviewers for their insight and comments that helped to improve this manuscript. Last, special thanks to the Bureau of Reclamation, Lower Colorado Region, for funding this publication.