Develop data science techniques for regional climate change studies

A grand challenge for Earth System science is to translate the influence of global processes that affect our climate into specific regional and local impacts. The research in CISL combines knowledge of Earth System models, downscaling methods, scientific workflows for large data sets, statistics, and the needs and constraints of local stakeholders. This effort therefore integrates CISL expertise in data science and impact assessment with the goal of transferring climate science into useful products for decision making in adaptation research and risk analysis.

In collaboration with a broad range of public and private laboratories and universities, the mission of the Regional Integrated Science Collective (RISC) is to generate high-quality regional-scale scenarios of projected climate change, make them widely available to the broader research community, and develop tools and methods for analyzing impacts, vulnerability, and adaptation options. RISC’s placement in IMAGe shows the close ties between evaluating climate models and quantifying uncertainty using statistics. RISC also reaches out to the broader decision-making and policy communities by integrating mathematical analyses into a more immediate and pragmatic realm. RISC has responsibility for serving large and multifaceted numerical experiments, so it is well aligned with CISL’s mission of data support to the climate science community.

RISC also develops and maintains the Weather and Climate Impacts Assessment Science Program (WCIASP) to investigate uncertainty in climate change research, study extreme weather and climate events, and support climate and health impacts workshops.

High-resolution research data for climate change studies

The Regional Integrated Science Collective (RISC) collaborates with public and private laboratories and universities to generate and evaluate regional-scale climate change scenarios using high-resolution regional climate models. Its goals are to:

  • Combine global and regional climate models to project regional climate change over North America and analyze resultant uncertainty, particularly focusing on process-level analysis.

  • Produce high-resolution, 150-year regional climate model simulations for North America.

CISL tackles the challenges of quantifying the uncertainty in predictions in ways that are useful for decision making and policy. This research supports CISL’s strategic goal 2: “Enhance the effective use of current and future computational systems by improving mathematical and computational methods for Earth System models and related observations.” It specifically fulfills CISL’s strategic imperative 2.1 to advance data-centric research by developing and applying novel data science techniques for regional climate change studies using high-resolution research data sets.

The research in RISC combines knowledge of Earth System models, downscaling methods, statistical techniques, and scientific workflows for large data sets with the needs and constraints of local stakeholders. This effort therefore integrates CISL expertise in data science and impact assessment with the goal of transferring climate science into useful products for decision making in adaptation research and risk analysis. Furthermore, guided by the NCAR strategic plan, CISL research improves predictions of weather and climate and estimations of their impacts.

NARCCAP and data products development for further analyses using NARCCAP: A centerpiece of RISC’s activity has been its leadership of the North American Regional Climate Change Assessment Program (NARCCAP). NARCCAP is systematically investigating the uncertainties in regional-scale projections of future climate. It is unique in its balanced design that allows for isolating the influence of individual regional and global models on the resultant climate simulations. The overall goal of NARCCAP has been to produce high-resolution (50 km) climate change scenarios using six regional climate models (RCMs) nested within four atmosphere-ocean general circulation models (AOGCMs) forced with the A2 SRES emission scenario, over a domain covering the conterminous U.S., northern Mexico, and most of Canada. The structure of the simulations is unique in its balanced design that allows for isolating the influence of individual regional and global models on the resultant climate simulations. The project also includes an evaluation arm whereby the participating RCMs were forced by reanalysis data sets. The resulting regional climate model runs and time slices formed the basis for multiple high-resolution climate scenarios that have been used in climate change impacts assessments in the U.S. and Canada. In FY2016, work related to NARCCAP has entailed further development of data products, detailed analysis of the simulations, and application of the data to numerous adaptation contexts.

RISC’s accomplishments in FY2016 include the development of a number of data products and services to support the users of NARCCAP data. These products will also be useful for future anticipated high-resolution regional climate simulations such as those being developed in North American CORDEX (see below). Seth McGinnis has been collaborating with CISL’s VETS section to help guide the development of next-generation data services that will enable users of output from Big Data projects like NARCCAP to access the data they need without downloading large volumes of unwanted data to get it. These new service capabilities include spatial and temporal subsetting, file spanning, aggregation, and format conversion.

Precipitation change predictions
These plots show the percent change in precipitation, future mid-21st century vs. current (1971-2000) period, from four simulations with the RegCM regional climate model. The upper left panel shows 50-km resolution driven by the MPI global model, the lower left panel shows the 25-km version. The upper right panel shows 50-km resolution driven by the GFDL model, and the lower right shows the 25-km version. Hatching indicates where the changes are statistically significant (at the 0.1 level). This image represents two types of uncertainty in future regional climate projections, based on different driving models and different simulation resolutions. Analyzing and comparing different types of uncertainty is important for more usefully and completely representing uncertainty in future climate.

Development of NA-CORDEX and preliminary analysis: The Co-ordinated Regional Climate Downscaling Experiment (CORDEX) has been ongoing for several years. It may be seen as an expansion of NARCCAP in the context of an international program. Regional models are driven by the more recent CMIP5 GCMS for a 150-year time period (through 2100), and the more recently developed representative concentration pathways (RCP) are used, including RCP8.5 and 4.5 The goal is to combine global and regional climate models to project regional climate change over North America and analyze resultant uncertainty, particularly focusing on process-level analysis.

Simulations continue to be performed even though North American CORDEX has been slow to advance through its schedule of simulations due to lack of sufficient funding from U.S. funding agencies. Linda Mearns of CISL and William Gutowski of Iowa State are NA-CORDEX co-chairs. In collaboration with Iowa State and the University of Arizona, simulations are being performed for a 150-year time period (1950-2100) over most of North America – approximately the same domain as NARCCAP – with two different regional climate models (RegCM4 and WRF), at two different spatial resolutions (50 km and 25 km) using ERA-Interim boundary conditions, and boundary conditions from three different GCMs that span the equilibrium climate sensitivity of the CMIP5 collection of global climate model simulations. An earlier version of RegCM (RegCM2) was NCAR’s original regional climate model.

Specifically, in RISC, in addition to the RegCM simulations produced last year at 25 and 50 km using the Max Planck Institute (MPI) global model boundary conditions for the RCP 8.5 concentration pathway, M. Bukovsky has produced similar simulations using WRF driven by the GFDL-ESM2M global model. Preliminary simulations with WRF driven by the HadGEM GCM have been produced, which will be completed in FY2017. When completed (including driving the RCMs by the HadGEM GCM) in early FY2017, a 2x2x3 matrix of simulations will be produced. Change in precipitation (current vs. mid-21st century) for the North American Monsoon region for the RegCM4 at the two different resolutions driven by the MPI and GFDL global models is presented in the figure above. Note that the different resolutions driven by the same GCM can result in different patterns of change, rather than producing the same pattern with greater detail in the higher-resolution case. The differences in the 50 km simulations driven by the two different global models are as great as the differences between the resolutions when RegCM4 was driven by the MPI global model. Other simulations have been produced by other groups, most notably in Canada using the Canadian regional climate model (CRCM5). The NA-CORDEX website lists all simulations being performed.

Archiving of NA-CORDEX data: Seth McGinnis and student assistant Daniel Korytina have made rapid progress in preparing and archiving subsets of NA-CORDEX data for all simulations completed so far. This includes archiving temperature and precipitation at multiple time scales (sub-daily to seasonal) for the entire NA-CORDEX domain. All data have been published.

Further development of simulation data bias correction for application to NARCCAP and NA-CORDEX data: Having identified bias correction as an important need of NARCCAP users – especially for impacts analysis – RISC has been working to bias-correct NARCCAP output using distribution mapping techniques. This work has broken new ground by applying the methods to daily data rather than monthly or seasonal climatologies. Seth McGinnis has developed a novel technique for distribution mapping called Kernel Density Distribution Mapping (KDDM). KDDM makes use of well-established statistical methods to perform distribution mapping using non-parametric estimates of the probability distributions underlying the data sets to be bias-corrected. This technique has been evaluated against existing techniques by use of an oracle analysis, wherein each technique is used to bias-correct synthetic data and the result is compared to a perfect correction, or “oracle.”

KDDM performs very well according to multiple metrics, and has the best performance on non-idealized data. It is also fast, robust, flexible, and conceptually straightforward. In FY2016, McGinnis improved on basic KDDM by refining the correction of extremes of daily temperature and precipitation. He evaluated the effect of bias correction using Kernel Density Distribution Mapping (KDDM) on extremes of precipitation and temperature. The KDDM methodology is effective at correcting the extremes of the data, and it produces better results than simply adjusting the mean and variance of the model outputs. In principle, it is possible to make additional small improvements to the correction of the tails using Extreme Value Theory, but in practice, this improvement is much smaller than the uncertainty associated with differences between the models, and does not in general warrant the extra effort involved. McGinnis is also developing an R package that implements the KDDM bias correction methodology. Application of KDDM to NA-CORDEX data will proceed in FY2017.

Process-level analysis of climate model results to determine the credibility of projections

Using the data generated by RISC for RISC scientific projects leads to greater understanding of the strengths and limitations of the data, which can then be communicated to the users so they can refine their investigations. Determining the credibility of climate projections is crucial for properly accounting for uncertainties in the projections.

This research supports CISL's strategic action items to develop and apply novel data science techniques for regional climate change studies using high-resolution research data sets, and to develop statistical methods to interpret geophysical data and improve model experiments.

Analysis of the southern Great Plains: RISC scientists and U. Connecticut colleagues produced a paper in J. Climate exploring the uncertainty in the NARCCAP data set for the Southern Great Plains region, and they established detailed credibility of the different climate change simulations. They determined that for this region there is consistency and credibility in all 12 simulations for changes in spring and summer precipitation. Warm-season precipitation will increase during the early spring wet season, but shift north earlier in the season, intensifying late summer drying. [Bukovsky et al., 2016: A Credible, Poleward Shift in Warm-Season Precipitation Projected for the U.S. Southern Great Plains. J. Climate (accepted, pending revision)].

Climate model bias analysis
Annual cycle of Snow Water Equivalent (SWE) averaged over North America from the six NARCCAP models (colors), the interquartile range of the observational ensemble data set (gray shading), and the median of the ensemble (black line). Units of area averaged SWE are in millimeters. Permanent ice points over Greenland and the northernmost Coast Range have been removed. This figure represents an advance in analyzing the climate model biases by taking into account the uncertainty in the observations, which can be quite large for SWE.

Analysis of snow water equivalent (SWE) in NCEP-driven NARCCAP simulations: Rachel McCrary, Seth McGinnis, and Linda Mearns have produced detailed analyses of the simulations by six different regional climate models. An additional part of this project entailed the development of an extensive set of observational data sets in order to include estimations of uncertainty in the observations to characterize in a more robust manner the biases in the regional models. The figure at right shows SWE aggregated over the full domain on a monthly basis for each model compared to the observations (including uncertainty in the observations). Note that the biases vary greatly from one model to another, for example, both CRCM and RCM greatly overestimate SWE, while WRFG substantially underestimates SWE. The paper describes how the biases in temperature, precipitation, and the nature of the surface package in each model largely explains the biases. [McCrary et al., Evaluation of Snow Water Equivalent in NARCCAP simulations, including measures of observational uncertainty. J. Hydrometeorology (submitted).]

Development, provision, and use of climate information including uncertainty measures for adaptation research

In collaboration with university partners, RISC scientists perform complex interdisciplinary research and work with stakeholders on projects relevant to climate change adaptation. This research is valuable in its own right, but it further serves to enrich the scientists’ understanding of stakeholder needs in the context of decision making under uncertainty. This understanding then feeds into their research on appropriately quantifying uncertainty in regional climate projections.

This work integrates CISL expertise in data science and impact assessment with the goal of transferring climate science into useful products for decision making in adaptation research and risk analysis. These efforts advance two of CISL’s strategic action items: to develop and apply novel data science techniques for regional climate change studies using high-resolution research data sets, and to develop statistical methods to interpret geophysical data and improve model experiments.

At end-FY2016, RISC has largely concluded work reported in FY2015 on three different research projects concerning adaptation to climate change at local and regional scales. In addition, work continued on one of two projects funded through the DoD Strategic Environmental Research and Development Program (SERDP): “Decision-Scaling: A Decision Framework for DoD Climate Risk Assessment and Adaptation Planning,” led by the University of Massachusetts. As part of this project Mearns, McGinnis, and colleagues at U. Massachusetts developed a novel approach to quantifying uncertainty in the climate change projections for the project by formulating priors for the Bayesian probabilistic model (joint PDFs for seasonal and annual temperature and precipitation) based on climate projections from the global climate models used to drive the NARCCAP models (Mearns et al., Characterization and Quantification of Uncertainty in the NARCCAP Regional Climate Model Ensemble and Application to Impacts on Water Systems, in preparation). When used in the context of a water resources model for the Colorado Springs area, the resulting PDFs (using either informed or uninformed priors) demonstrated that, even at the annual time scale, important differences resulted in the water resource reliability domain covered by the PDFs. Using the informed priors resulted in a smaller percentage of the joint PDF falling in the unreliable portion of the water resources reliability domain.

WCIASP

RISC also maintains and develops the Weather and Climate Impacts Assessment Science Program (WCIASP). The goal of WCIASP is to improve society’s ability to manage weather and climate risks by creating and providing research tools and methods at the critical frontiers of impact assessment science. WCIASP has three primary thrusts: investigating uncertainty in climate change research, studying extreme weather and climate events and their impacts, and supporting the Climate and Health Workshop series. In FY2016 WCIASP funded projects throughout NCAR, particularly in RAL and IMAGe.

This work supports CISL’s strategic imperative to advance data-centric research by combining climate data products and impact projections with companion measures of uncertainty. It also supports CISL’s imperative to foster research collaborations by drawing scientific, mathematical, and computer science researchers to NCAR and engaging them in CISL’s research efforts.

Projects funded in FY2016 included four in RAL and two in IMAGe/RISC: development and maintenance of the Extremes Toolkit, uncertainty in vulnerability assessments, integrated uncertainty in water resource assessments, planning for the climate and health workshop, NA-CORDEX regional climate model simulations and data archiving, and analysis of snow in the NCEP-driven NARCCAP models.

Funding sources

The Regional Integrated Science Collective (RISC) and Weather and Climate Impacts Assessment Science Program (WCIASP) are primarily supported by NSF Core funding as well as interagency support for NARCCAP and the use of NARCCAP results for adaptation planning from NSF, NOAA, and the U.S. Department of Defense Strategic Environmental Research and Development Program (SERDP).