Methods

Study domain and temporal extent

The dataset represented here provides projections of streamflow at 396 sites throughout the Pacific Northwest. Most of the sites are located within the Columbia River basin but the dataset also includes selected sites within the coastal drainages in Washington and Oregon.

The hydrological models were implemented at a spatial resolution of 1/16º, which corresponds to roughly rectangular grid cells that are 6 km or 3.5 mi on a side, resulting in 23,929 individual hydrological model elements (or grid cells) over the domain.

Most of the climate change simulations were run from water year 1950 through 2099. The exceptions are (1) the historical simulation which runs from 1960 through 2011; and (2) the time series resulting from the dynamically-downscaled meteorological forcings produced by collaborators at Oak Ridge National Laboratory.

Products

The main product of this project consists of streamflow time series for 396 locations. Because of the different methodological choices, there are 172 individual time series for each of these locations. In addition, for a subset (190) of these locations, we provide both raw and bias-corrected versions of these time series. All told, the streamflow dataset includes 100,792 individual streamflow time series. While this number may be overwhelming in aggregate, many users may only use a selected subset or slice from that large ensemble. We provide separate streamflow time series for every location in the domain for each climate scenario, global climate model, downscaling method, and hydrologic model set-up.

The River Management Joing Operating Committee released a report on the general findings of the study, which provides a synopsis of methods as well as results for different regions around the Pacific Northwest.

Methodological choices

One purpose of this study is to evaluate the impact that methodological choices have on streamflow projections under climate change. To construct a dataset that could probe that question, we devised a modeling tree which involved four decision points where we could choose among multiple options. These methodological choices were made with respect to:

  • Representative concentration pathway
  • Global climate model
  • Downscaling method
  • Hydrological model

Representative concentration pathway (RCP)

The Representative concentration pathways (RCPs) describe four different 21st century pathways of greenhouse gas emissions and atmospheric concentrations, air pollutant emissions and land use (IPCC 2014). The RCPs are described in detail by Vuuren et al. (2011) and are available online. In turn, these RCPs were used by the Coupled Model Intercomparison Project Phase 5 (CMIP5) to create climate projections associated with each RCP.

As explained by IPCC (2014)), “The RCPs include a stringent mitigation scenario (RCP 2.6), two intermediate scenarios (RCP 4.5 and RCP 6.0) and one scenario with very high GHG emissions (RCP 8.5). Scenarios without additional efforts to constrain emissions (’baseline scenarios’) lead to pathways ranging between RCP6.0 and RCP8.5 (Figure SPM.5a). RCP2.6 is representative of a scenario that aims to keep global warming likely below 2°C above pre-industrial temperatures.” In this project, we elected to evaluate climate projections associated with the lower intermediate scenarios and the highest scenario: RCP 4.5 and RCP 8.5. For the study domain, both scenarios produce a warmer future than historically observed, with RCP8.5 warmer than RCP4.5. It is important to keep in mind that the RCPs are possible concentration pathways and that actual greenhouse gas emissions may be lower or higher. The projected streamflow time series in this dataset are predicated on our choice of RCP 4.5 and RCP 8.5.

Global climate model (GCM)

Global climate models (GCMs) use the RCPs to supply boundary conditions to simulate different realizations of the future of the Earth’s climate. Outputs from ten different global climate models were used in this study to investigate the uncertainty within hydrologic projections owing to choice of GCM. The CMIP5 experiment involved 31 GCMs, but this project uses outputs from 10 of them. The models were chosen based upon a variety of metrics detailed in Rupp et al (2013). While the earlier Columbia River climate change project (Hamlet, 2013) used GCMs from many of the same modeling groups, these models have been significantly enhanced and updated.

Downscaling method

Downscaling is a technique which translates meteorological information at a relatively coarse spatial scale (~150 km) to the scale of the hydrologic model implementation (~6 km). Downscaling methods often incorporate a bias-correction step which removes systematic biases that the GCMs have in their simulations of the 20th century. For example, a particular GCM may be wetter and warmer than observed climatology when compared for a historical period. In climate change studies, we are predominantly concerned with the change in climate over a period of time. Therefore, we are interested in the change signal and correct the model output for historic biases. After all, if we use the GCM output that is too wet and too warm as input to a hydrology model, then we may not be able to simulate hydrological processes such as snow accumulation and melt with sufficient realism.

Three different downscaling/bias-correction methods were used to create the meteorological data for this project:

  • The bias-correction, spatial-disaggregation (BCSD) technique (Wood et al, 2004). This method uses a monthly quantile mapping approach to remove systematic biases at the GCM grid scale. It then uses an inverse square weighting method to map the GCM outputs onto the fine scale spatial variability of the domain. These meteorological forcings are available for 1950-2099. This is the method which was used in the RMJOC-I project.

  • The multivariate adaptive constructed analogs (MACA) technique (Abatzoglou and Brown, 2012). This method was developed and implemented by John Abatzoglou and Katherine Hegewisch at the University of Idaho. Like BCSD, it uses a training historical dataset to remove GCM biases. However, unlike BCSD, MACA uses an analog approach to match spatial patterns in global climate model output to map them to the fine scale variability of the domain. Another key difference is that the MACA approach uses daily output from the GCMs, while BCSD uses only monthly output. Like the BCSD forcings, these are available from 1950-2099.

  • For a subset of GCMs and scenarios, we were able to use a set of dynamically-downscaled outputs developed by Moetasim Ashfaq and Shih-Chieh Kao at Oak Ridge National Laboratory (Ashfaq et al., 2016). This method involved running a regional climate model to dynamically downscale the GCM output to an 18 km resolution. This output was then statistically bias-corrected to the resolution that is used as input for the hydrologic model. It is important to note that these forcings were trained to a different historical meteorological forcing dataset. Further, unlike the two statistical techniques described above, these forcings are only available from January 1966-November 2005 and January 2010-November 2050.

Hydrologic models

Four different hydrologic models were used in the development of the dataset. Three were distinct implementations of the Variable Infiltration Capacity (VIC) model and a fourth was an implementation of the Precipitation Runoff Modeling System (PRMS). The VIC model used in this project involved a novel implementation of a simple glacier model. In the naming conventions of the dataset, the four different model set-ups are denoted by the hydrologic model followed by “PX” where X=1,2,3.

  • VIC-P1: This implementation of VIC uses parameters calibrated by the University of Washington to the NRNI flow dataset provided by the RMJOC.

  • VIC-P2: This implementation of VIC uses parameters calibrated by collaborators at Oak Ridge National Laboratory (Naz et al., 2016).

  • VIC-P3: This implementation of VIC uses parameters calibrated by collaborators at the Research Applications Laboratory at the National Center for Atmospheric Research.

  • PRMS-P1: This implementation of PRMS uses parameters calibrated by the University of Washington to the NRNI flow dataset provided by the RMJOC.

Routing model

The streamflow from all hydrologic model setups was routed through a stream network using the same routing model and set-up. The RVIC model is a source-to-sink model developed originally by Lohmann et al. with improvements by (Hamman et al. 2017). See rvic.readthedocs.io for detailed documentation on the model.

Streamflow bias-correction

At 190 sites throughout the basin, the RMJOC provided time series of no-regulation, no-irrigation (NRNI) streamflow. These reference streamflows were used, where available, to adjust the modeled streamflow time series to remove systematic biases. The method uses the preservation of ratio technique described by Pierce et al. (2015), with a number of adaptations to ensure that any change in annual volumes (i.e. a 15% increase or decrease in overall streamflow at a given site) is preserved between the raw and bias-corrected time-series. It is important to note that, as with most streamflow bias-correction procedures, the method breaks the water balance of the hydrologic modeling system, adding or removing water from the system such that there can be discontinuities between upstream and downstream locations. In addition, all flow locations are bias corrected independently, which may result in inconsistencies between upstream and downstream locations at short time scales.