Colorado River water supply is predictable on multi-year timescales owing to long-term ocean memory

Colorado River water supply is predictable on multi-year timescales owing to long-term ocean memory


Skillful multi-year climate forecasts provide crucial information for decision-makers and resource managers to mitigate water scarcity, yet such forecasts remain challenging due to unpredictable weather noise and the lack of dynamical model capability. Here we demonstrate that the annual water supply of the Colorado River is predictable up to several years in advance by a drift-free decadal climate prediction system using a fully coupled climate model. Observational analyses and model experiments show that prolonged shortages of water supply in the Colorado River are significantly linked to sea surface temperature precursors including tropical Pacific cooling, North Pacific warming, and southern tropical Atlantic warming. In the Colorado River basin, the water deficits can reduce crop yield and increase wildfire potential. Thus, a multi-year prediction of severe water shortages in the Colorado River basin could be useful as an early indicator of subsequent agricultural loss and wildfire risk.


The Colorado River is the most important water resource in the semi-arid western United States (U.S.). Demand for its water has increased continuously for the past 60 years but its water supply is facing an unsustainable future due to declining precipitation and prolonged drought events1,2,3. Prolonged water shortages in the Colorado River can cause serious damages to a wide range of sectors, including agriculture, forestry, energy, food security, drinking water, and tourism4,5. High-impact drought events in the Colorado River basin during 2000–2016 severely stressed regional water supply and strained reservoir operations, recreation, and ecological services1,3. To help managers and policy-makers cope with drought-induced water shortages coupled with ever-increasing demand, it is crucial to develop multi-year drought predictions for safeguarding and maintaining industrial and societal wellbeing.

Presently, U.S. operational drought forecasts primarily focus on day-to-month outlooks, such as the short-term drought severity indicator and the monthly drought outlook generated by the Climate Prediction Center (CPC) and the National Drought Mitigation Center of the National Oceanic and Atmospheric Administration. These forecasts are severely limited by short-term weather phenomena6, since unpredictable atmospheric noise makes interannual-to-decadal climate prediction challenging7,8. However, previous research has identified a prominent “drought cycle” in the Intermountain West and the southwest U.S., which was attributed to climate variability associated with long-term ocean memory9,10,11,12,13,14. Besides ocean memory, land systems (i.e., soils, groundwater, streamflow, vegetation, and perennial snowpack) filter out the high-frequency precipitation fluctuation and integrate atmospheric signals over space and time15,16,17,18,19. These results lead to a hypothesis proposed herein that skillful multi-year predictions of the Colorado River water supply are possible by utilizing long-term ocean memory, its associated atmospheric teleconnections, and the natural filtering effect in the land system altogether. A similar concept has been applied to develop seasonal drought forecasts based on El Niño Southern Oscillation (ENSO) predictions20 and a statistical model for multi-year water supply predictions21,22, yet its application to the water supply forecast beyond seasonal timescales remains unknown due to the signal-to-noise paradox7.

Here we demonstrate that interannual-to-decadal variability of the Colorado River water supply is predictable for several years in advance using a decadal climate prediction approach. Our assessment is based on three experiments using the fully coupled climate model Community Earth System Model (CESM; see “Methods”): a historical and future emission scenario simulation (the externally forced run), an ocean data assimilation run (referred to as the ASSIM run), and a multi-year initialized prediction experiment (the hindcast run). In the ASSIM run, we assimilated the observation-based 3-dimensional ocean temperature and salinity anomalies23 into the ocean component of CESM, with prescribed natural and anthropogenic radiative forcings. As a result, the model-simulated atmosphere-land variability in the ASSIM run is indicative of the response to ocean variability and external radiative forcings. We conducted the ocean assimilation experiment for the period 1960–2015 with 10 ensemble members. To minimize the bias known as artificial model drift during the simulation, an effective bias-adjustment method was adopted and applied in the assimilation method24. Predictability of Colorado River water supply and its sources are determined by the externally forced run and the hindcast run.


Reconstruction of Colorado River water supply

To detect water shortages of the Colorado River, we used a historical record of water supply in the Colorado River basin provided by the Bureau of Reclamation1 (blue line in Fig. 1a). This Colorado River water supply dataset was designed to develop adaptation and mitigation strategies for water resource agencies and stakeholders throughout the Colorado River basin. Its interannual-to-decadal variability mostly reflects the natural flow at Lees Ferry, Arizona, and is assumed to be free of human water usage (e.g., irrigation)25. This water supply record illustrates severe shortages in the years 1963, 1977, 1981, 1990, 2002, 2012, and 2013 (yellow shade in Fig. 1a). These severe shortages correspond with agricultural losses and high fire activity in the Colorado River basin, as described later.

Fig. 1: Observed and reconstructed Colorado River water supply.



a  Annual mean time series of Colorado River water supply (blue) and its reconstructions by CPC soil moisture reanalysis (black broken) and the ensemble mean of 10-member ASSIM runs (black solid). The reconstructions of the Colorado River water supply are obtained by area averaged soil moisture anomalies in CPC reanalysis (33°N–43°N, 119°W–103°W) and ASSIM runs (black boxes in  d e ). Years with prominent water supply shortages are highlighted with yellow shading (1963, 1977, 1981, 1990, 2002, 2012, and 2013). Correlation maps of  b c  precipitation and  e f  soil moisture anomalies in the observation-based estimate (left) and the ASSIM run (right) against the Colorado River water supply for 1960–2015. Red line and yellow mark in  b d  indicate geographical locations of Colorado River and Lees Ferry, respectively. Correlation coefficients of 0.28, 0.33, 0.45, and 0.53 correspond to the statistical significance at 90%, 95%, 99%, and 99.9% levels with 36 degrees of freedom on the basis of two-sided Student’s  t -tests and an equivalent sampling size of the Colorado River water supply. Anomalies are defined as deviations from climatological means and linear trends are removed in each grid.

Full size image

According to previous studies, the major fraction of Colorado River streamflow variability arises from groundwater variability through precipitation input as well as land processes of infiltration, subsurface storage and transmission, and convergence toward the channels17,26,27. Through these regional processes, Colorado River streamflow is considered to have a large spatial footprint of hydroclimate conditions in the Intermountain West17. To reveal regional climatic control of the Colorado River water supply, we made correlation maps of precipitation and total soil water (i.e., the sum of soil water in all soil layers) as observation-based and model-simulated anomalies with the observed Colorado River water supply (Fig. 1b–e). While the correlation between the Colorado River water supply and annual mean precipitation variability is high, we find more significant correlations of water supply with soil water anomalies across the western U.S. (Fig. 1b, d). The ASSIM run also demonstrates that the Colorado River water supply positively correlates with soil water anomalies around Utah, Colorado, and New Mexico (Fig. 1e), albeit with somewhat distorted spatial patterns of precipitation and soil water anomalies compared to the observation. By taking the regional average over these highly correlated soil water anomalies (black boxes in Fig. 1b, e), we reconstruct the Colorado River water supply using the National Centers for Environmental Prediction CPC reanalysis and the ASSIM run (black lines in Fig. 1a). This reconstructed water supply shows a close agreement with the observed temporal variation of the Colorado River water supply. This result supports our hypothesis that the Colorado River water supply is closely linked to soil water variability in the Intermountain West.

Atmospheric weather disturbances still generate unpredictable high-frequency noise, inhibiting precipitation predictability, yet such high-frequency components of precipitation variability are mostly filtered out through the land hydrological processes in the Intermountain West11,15. Because of this land filtering effect, annual soil water anomalies reflect precipitation anomalies averaged for several years (Supplementary Fig. 1) that link with long-term ocean memories. Consequently, we reconstructed the Colorado River water supply from the areal average of soil water anomalies associated with interannual-to-decadal ocean variability in the ASSIM run. The 10-member ensemble mean of ASSIM runs demonstrates skill in capturing many major historical shortages of the Colorado River water supply (Fig. 1a), even though it did not include any atmospheric or land observations. The correlation coefficient between the observed and the model reconstructed Colorado River water supply was higher when we applied a 3-year running mean filter to highlight the low-frequency climate variability (correlation coefficients are 0.43 for annual and 0.60 for 3-year means, respectively). Our analysis suggests that the Colorado River water supply is potentially predictable through “perfect knowledge” of ocean conditions.

Predictability of Colorado River water supply

Next, we examined the dynamical retrospective forecasts for the period 1960–2015, which was initialized annually on January 1st based on the ASSIM run (referred to as the hindcast run; see Methods). This hindcast experiment consists of 10-year-long predictions with 10 ensemble members for each initialized run. The ensemble mean of the hindcast runs for the reconstructed Colorado River water supply and the corresponding ensemble spread aligned well with the CPC reanalysis and the ASSIM run for 1-year and even 2-year lead times (Fig. 2a, c). The predictive skills measured by the anomaly correlation coefficient (ACC) and the root-mean-square error (RMSE) in our hindcast run outperform those in the externally forced run and persistent forecast (see “Methods”), up to the lead time of 48 months (Fig. 2b, d). These higher skills in the hindcast run indicate that ocean initialization is crucial for skillful forecasts of the Colorado River water supply on interannual-to-decadal timescales whereas atmospheric initialization can improve land hydrological predictability on seasonal timescales28. As a result, the hindcast run demonstrates skill in predicting the Colorado River water supply for 2 years in advance (Fig. 2b, d). Furthermore, we find higher predictive skill by measuring the hindcast run against the model reconstruction of the ASSIM run (i.e., red dotted lines in Fig. 2b, d) compared to either the observation-based products of CPC reanalysis or the observed Colorado River water supply. Because of deficiencies in current climate models simulating observed land hydrology, these results suggest that enhanced predictive skill could be achieved by improving model performance in simulating soil water variability. The dynamical prediction presented here supports earlier statistical predictions of Colorado River water supply years ahead21,22,29,30.