Open data and tools for multiple methods of global climate downscaling

Climate change will impact nearly every sector of society. Governments, advocacy groups, and corporations are increasingly confronting the need for robust estimates of climate-related risks and uncertainties. Planners and regulators are incorporating future climate information into their decision-making. The United States, for example, conducts a National Climate Assessment (NCA) every four years which informs its Climate Adaptation and Resilience Plans, and the Securities and Exchange Commission has proposed mandating public disclosure of climate-related risk. Internationally, the Intergovernmental Panel on Climate Change (IPCC) releases global assessments of projected future impacts of climate change on human, physical, and ecological systems.

Climate impact and risk assessments rely critically on global climate models (GCMs). However, projections from GCMs in raw form are rarely directly usable in impact assessments, because, for example, they are too coarse in spatial scale or contain persistent biases. A process called “downscaling” produces datasets that are spatially and temporally tuned for detailed projection of climate impacts and risks. There are a variety of downscaling approaches, each involving complex analytical workflows, and the choice of downscaling algorithm can strongly influence the resulting climate projections.¹1T Lopez-Cantu et al. (2020) Uncertainties in Future U.S. Extreme Precipitation From Downscaled Climate Projections Geophysical Research Letters

We are releasing a new global downscaled climate data product spanning multiple downscaling methods, alongside an interactive mapping tool for inspecting, exploring, and comparing the data. Our release includes four downscaling algorithms run globally for the first time on the latest generation of climate projections from the Coupled Model Intercomparison Project Phase 6 (CMIP6), which forms the basis of reports from the IPCC. Our data product adds to a growing collection of downscaling algorithms and datasets, including recent releases from NASA-NEX and the Climate Impact Lab. Together with our fully open source analyses and map tool, we hope this release enables important comparisons and evaluation of this critical step in climate impacts analysis.

In this article, we explain what downscaling is and why it matters, the key contributions of our first data release, and how we expect the data and tools to be used.

The role of downscaling

Projections of future climate are often derived using global climate models (GCMs)²2GCMs are increasingly referred to as Earth System Models (ESMs), though we use the term GCM in this article.. GCMs are scientific models that simulate the earth’s climate as a whole, including oceans, atmosphere, and land. They are intended to capture continental scale climatological patterns, and are shown to capture broad patterns from meteorological observational datasets of the 20th century. These GCMs provide well-tuned sandboxes for testing the effect of different future greenhouse gas emissions on the planet’s climate.

By construction, GCMs subdivide the planet into pixels and yield a projection of the future climate for each pixel on every day. Depending on the GCM, a pixel might be ~1-2 deg on a side, or about 100-200 km at the equator.

There are two key barriers to the immediate use of raw GCM data within a risk estimation framework. First, the spatial resolution is too coarse for direct use in many applications. For example, a 100 km pixel in a GCM may be assigned a single elevation, when it actually includes a mountain range or coastline. Temperature can vary dramatically across these topographies which the GCMs’ coarse scale would miss. Precipitation patterns are similarly influenced by such topographic patterns.

Second, the GCMs are designed to capture global patterns, but they are not expected to reproduce observed climate patterns in any particular location. As a result, they will perform better in some locations and worse in others, and the magnitude and direction of any “biases” will differ across GCMs. This issue matters especially in downstream applications that are sensitive to specific thresholds. For example, predicting snowpack requires knowing when temperatures are above or below freezing (0 degrees celsius). If the temperature at a given location hovers around 0 degrees, a small bias could dramatically impact predictions. Similarly, GCMs have long struggled to accurately capture the number of wet days each year, a statistic that is important in many agricultural and hydrologic applications.³3A Dai et al. (2006) Precipitation Characteristics in Eighteen Coupled Climate Models Journal of Climate

These barriers do not make GCMs useless. On the contrary, global climate simulations contain invaluable information for predicting the planet’s future response to elevated greenhouse gasses. However, to use that information for more granular impact assessment, bias-corrected higher resolution data is needed.

Scientists employ a process called “downscaling” to translate coarse resolution climate projections into higher resolution climate data at a scale that resolves the most influential climatic variations and removes biases with respect to historical observations. In general terms, statistical downscaling involves “training” a model on the statistical relationship between a GCM and an observational dataset over a historical period, and then applying that model to generate future predictions.⁴4Statistical downscaling is distinct from “dynamical” downscaling, which instead involves running a numerical weather model to derive climate variables. The downscaled data can then provide an input into downstream impact models and risk assessment (Figure 2).

While the broad goals of downscaling are easy to define, there are many different methods with distinct strengths which can be matched to the desired application of the downscaled data. Downscaling methods include several steps (see Table 1) where influential choices need to be made, including the form of bias correction, whether to use one or multiple predictor variables, and the incorporation of regional climate patterns (see “Different algorithms” below). All of these decisions can affect the resulting downscaled dataset, and the resulting differences propagate through follow-on impacts models, creating uncertainty around projections. For example, this cascading uncertainty has been explored extensively within the context of water resources.⁵5M S Khan et al. (2006) Uncertainty analysis of statistical downscaling methods Journal of Hydrology ^,66J Chen et al. (2013) Performance and uncertainty evaluation of empirical downscaling methods in quantifying the climate change impacts on hydrology over two North American river basins Journal of Hydrology ^,77A Wootten et al. (2017) Characterizing Sources of Uncertainty from Global Climate Models and Downscaling Techniques Journal of Applied Meteorology and Climatology ^,-88O S Chegwidden et al. (2019) How Do Modeling Decisions Affect the Spread Among Hydrologic Climate Change Projections? Exploring a Large Ensemble of Simulations Across a Diversity of Hydroclimates Earth's Future Differences between methods are often most pronounced when projecting climatic extremes, such as intense precipitation¹1T Lopez-Cantu et al. (2020) Uncertainties in Future U.S. Extreme Precipitation From Downscaled Climate Projections Geophysical Research Letters or heat waves.⁷7A Wootten et al. (2017) Characterizing Sources of Uncertainty from Global Climate Models and Downscaling Techniques Journal of Applied Meteorology and Climatology Differences between methods can be further magnified when downscaled datasets are ingested into downstream impacts models. It is thus critical for the climate impacts community to be able to access a variety of downscaled datasets and to compare their suitability for a desired application.

What we’re releasing

We are releasing a new collection of downscaled datasets produced from multiple statistical downscaling methods, alongside a companion map tool to explore the data. We used climate change simulations from up to six global climate models from CMIP6 and three different future emissions scenarios, and we applied four different downscaling methods: MACA⁹9J T Abatzoglou et al. (2011) A comparison of statistical downscaling methods suited for wildfire applications International Journal of Climatology , DeepSD,¹⁰10T Vandel et al. (2017) DeepSD: Generating High Resolution Climate Change Projections through Single Image Super-Resolution Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining and two implementations of GARD¹¹11E D Gutmann et al., in review (Table 1). We chose to run a modest number of GCMs and scenarios, and instead focused on developing, testing, and implementing multiple downscaling methods. There is no one correct downscaling method, and we do not suggest any of our implementations is an optimal dataset. Rather, we hope our focus on methodological diversity complements existing efforts, and will help the research and applications communities evaluate uncertainties associated with downscaled data products.

Our datasets include projections of daily maximum and minimum temperature and precipitation through the end of the 21st century for the entire globe at a 0.25 degree resolution (approximately 25 km x 25 km).¹²12The MACA method was not implemented over Antarctica We trained all models on the ERA5 reanalysis product.¹³13H Hersbach et al. (2020) The ERA5 global reanalysis Quaterly Journal of the Royal Meteorlogical Society In the theme of open data development, we welcome feedback from the scientific community so as to guide future improvements. We anticipate releasing updates to expand or refine the collection of datasets.

Because the data are quite large — multiple terabytes for the complete collection — we also built a new interactive mapping tool that makes them explorable and comparable directly in a web browser, hopefully lowering the barrier of entry for those interested in exploring or working with the data. The map tool shows downscaled data alongside the raw (coarse) CMIP6 data. We expect this tool to be particularly useful for orienting users in space and time, facilitating model and variable selection, and accelerating quality control within and across models and downscaling methods. We also hope it enables a wider audience to explore patterns in the data and identify opportunities for future improvement.

Below, we highlight three of our key contributions in more detail.

01 — Different algorithms

As discussed above, different algorithms produce different downscaled climate change projections (Figure 3). To date, the uncertainty associated with choice of downscaling methods has received less attention than other physical modeling steps in climate impacts (e.g. uncertainties associated with GCM selection). Systematic comparisons have been made for downscaled CMIP5 data products,¹⁴14M A Sunyer et al. (2015) Inter-comparison of statistical downscaling methods for projection of extreme precipitation in Europe Hydrology and Earth System Sciences ^,1515A T Werner et al. (2016) Hydrologic extremes – an intercomparison of multiple gridded statistical downscaling methods Hydrology and Earth System Sciences ^,1616D Mullan et al. (2015) Validation of non-stationary precipitation series for site-specific impact assessment: comparison of two statistical downscaling techniques Climate Dynamics ^,-1717H Tabari et al. (2020) Comparison of statistical downscaling methods for climate change impact analysis on precipitation-driven drought Hydrology and Earth System Sciences but most comparisons have been local or regional in scale, and the community has not yet been able to assess downscaling sensitivities for CMIP6. To help support such comparisons, we focused on implementing multiple algorithms (Table 1), which differ in three important ways.

First, methods differ in the variables they use for training and inference. For example, BCSD¹⁸18A W Wood et al. (2004) Hydrologic Implications of Dynamical and Statistical Approaches to Downscaling Climate Model Outputs Climatic Change uses a single-variate approach, in which the model predicts the same variable (e.g. precipitation) on which it was trained. In contrast, the Ensemble Generalized Analog Regression Downscaling (En-GARD) model can use any number of time-varying climatological or biogeophysical variables. We implemented¹⁹19We refer to our implementation as “GARD” because we have not included the Ensemble element of the mode two variants of GARD: (1) a single-variate implementation and (2) a multi-variate implementation which uses zonal and meridional winds along with the target variable as predictors. As yet another example, DeepSD method can incorporate additional variables, like elevation, as time-invariant features.

Second, downscaling methods incorporate varying levels of spatial information. Some methods, like BCSD, only use local GCM information, whereas others use larger regional patterns. MACA, for example, constructs a final, fine-scale product informed by coarse-scale weather patterns from the reference dataset. Similarly, DeepSD learns generalized spatial patterns based on observational and topographic data using a computer vision approach. The consequences of these choices for specific applications will require extensive analysis and comparison.

Third, methods range in their maturity, from established to experimental. BCSD was introduced in the early 2000s and has been used in a range of local and global applications; MACA was introduced 10 years later and has also become well-accepted due to some notable improvements around spatial patterns (discussed above) and extreme events. The other methods, especially DeepSD, are more recent and experimental and require more development and testing for real-world application.

02 — Transparency

The downscaled datasets we are releasing here are accompanied by the open source code that we used to produce them. This transparency is important for two reasons: First, it allows a broader community to evaluate the results and provide feedback. Second, it enables a better understanding of methodological assumptions and caveats that may be relevant when choosing data for climate impact assessment. We hope this work can be used as an example of how large-scale climate datasets can be developed in the open: contributing to open source packages and developing within open codebases, and creating publicly available datasets fully inspectable in a web browser.

03 — Information equity

Making downscaled climate data globally equitable requires both having comprehensive coverage across regions and methods, and ensuring data is as accessible as possible. Historically, downscaled data has been most readily available in North America and Europe, and outside of those regions there are especially few examples of multiple downscaled datasets available for comparison. Alongside other recent efforts, we hope that the robust global coverage and multiple methods in this release help enable global and regional applications that make use of downscaled climate data. Regarding accessibility, we recognize that working with these data, let alone producing them, still demands significant computing resources. Through our map tool, we hope to lower the barrier to basic exploration and evaluation of the data, but recognize that more complex use cases continue to require significant technical capacity (see Table 2).

Comparisons to existing work

There is a growing landscape of CMIP6 climate data sources, and we hope that our release both complements these data products and helps facilitate easy comparisons among them. Our four downscaled datasets join two other releases of global CMIP6 downscaled data: a BCSD implementation from NASA Earth Exchange (NEX) Global Daily Downscaled Projections and a quantile-mapping based approach²⁰20A J Cannon et al. (2015) Bias Correction of GCM Precipitation by Quantile Mapping: How Well Do Methods Preserve Changes in Quantiles and Extremes? Journal of Climate conducted by the Climate Impact Lab. The NEX dataset notably includes additional variables not included in either this release or the CIL release, including relative humidity, specific humidity, surface wind, and downwelling shortwave and longwave radiation. The NEX and CIL datasets also include a larger set of GCMs and scenarios than we have released here. Given that all of these datasets are now publicly accessible, there is an opportunity to systematically compare them. We note, however, that the NEX product used a different training dataset — GMFD rather than ERA5 — so comparing it to the other downscaled datasets reflects both the difference in algorithm and the difference in training data, which can have a substantial impact on both historical and future projections.

Thus far we have only considered “statistical” downscaling methods, but efforts are also underway to develop dynamically-downscaled data (e.g. the CORDEX initiative). That approach relies on using GCM outputs as boundary conditions to a higher resolution weather model and involves an entirely different set of modeling decisions than what we have discussed above.

How these data can be used

In the following table, we detail how you might use these data and tools depending on your goals.

What comes next

We hope that this release furthers the ability of the climate impacts community to consider the consequences associated with choosing downscaling methods or downscaled datasets. We particularly acknowledge researchers in domains dependent on extreme events, where these choices are especially significant.

Looking ahead, we are eager for others to explore these downscaling implementations and datasets. We anticipate expanding the collection of datasets along several dimensions — additional downscaling methods, more GCMs and scenarios, alternative observational datasets — pending community feedback. We plan to continue exploring how to make these data accessible to all, and finding ways to leverage them in new applications.