Downscaling

The goal of downscaling is to create locally accurate climate information from global-scale data by placing it in the context of observed local climatological conditions.

This improves the spatial and temporal resolution of the data, making it more useful for local and regional analyses.

Downscaling results. LEFT: Original CRU data at 0.5 x 0.5 degrees. RIGHT: the same CRU data downscaled to 2 x 2 km.

Dynamical downscaling

Uses a physically based weather forecasting model to produce higher time and space resolution data from coarser General Circulation Model (GCM) data.

Dynamical downscaling details: the WRF model

Dynamical downscaling uses a physically based weather forecasting model to produce higher time and space resolution data from coarser General Circulation Model (GCM) data.

SNAP uses the Weather Research and Forecasting (WRF) model to downscale GCM data. The internal physics engine of the WRF model is bounded by the input data. For projected data, it is bounded by GCM output, and for historical periods, it is bounded by reanalysis data.

Fewer models and scenarios are available because this is a very computationally expensive process. However, more than 50 variables are available at an hourly temporal resolution. We also provide daily and monthly versions. Spatial resolution is 20-km x 20-km. Example variables include:

Heat fluxes (radiative and turbulent)
Precipitation type (convective and non-convective)
Rainfall
Snow depth
Snowfall
Soil temperature and moisture
Temperature
Upper level winds, heights, and temperatures
Wind speed and direction

See Bieniek et al. (2016) for a more detailed description of the WRF model procedure and an evaluation against historical temperature and precipitation data.

Statistical (delta) downscaling

Adds the difference (delta) between a historical period and a modeled value to a known historical climatology.

Statistical downscaling: the delta method Why a proportional anomaly for precipitation? References

Statistical downscaling: the delta method

The statistical method we use is often called the delta method, as it adds the difference (delta) between a historical period and a modeled value to a known historical climatology. Thus, only the change in climate (as expressed in the model) is incorporated into a known historical baseline.

This method uses only subtraction and division, which helps in interpreting and explaining downscaling results
Its low computational demand makes it very efficient to downscale many GCMs and emission scenarios over hundreds of years
Monthly temporal resolution
Spatial resolution is 2-km x 2-km pixels across Alaska and Western Canada
Variables are limited those that have long observational records, including surface temperature and precipitation.

General steps

Determine which time series you want to downscale and which climatology dataset you will downscale to. We downscale GCM and CRU time series, and use PRISM and CRU climatologies as downscaling climatologies.
Calculate changes in the monthly time series (temperature or precipitation) in relation to the time series average climate during the time period for which the climatology is available.
Interpolate those changes—also referred to as deltas or anomalies—to match the climatology’s spatial resolution. Then, add them to (for temperature) or multiply them by (for precipitation) the climatology values for the same month.

SNAP delta downscaling code can be obtained from our code repository.

Example: downscaling GCM data to PRISM climatology

Standardize GCM and PRISM data
- Rotate grid and set latitude and longitude values to standard WGS84 geographic coordinate system. This sets North as the top of the grid and converts original lat/long values from 0°-360° to -180° -180°.
- Convert GCM units to more common units:
  - Convert temperature from Kelvin to Celsius
  - Convert precipitation values from kg/m²/sec to mm/month
Calculate GCM climatologies to determine projected changes in climate and the amount of model bias inherent in that change. Climatologies are GCM mean monthly values across a reference period (usually 30 years) from the 20c3m scenario outputs. The values represent modeled data and contain an expected model bias that we later adjust.
- Determine a reference state of the climate according to the GCMs by using 20th-century (20c3m) scenario GCM data values to calculate climatologies for the same time span used in the high resolution data we are downscaling to.
- We do this calculation for a worldwide extent at the coarse GCM spatial resolution (range: 1.875°–3.75°).
Calculate anomalies: monthly absolute (for temperature) or proportional (for precipitation) anomalies for a worldwide extent at the coarse GCM spatial resolution.
- Temperature: future monthly value (e.g., May 2050 A1B scenario) - 20c3m climatology
- Precipitation: future monthly value (e.g., May 2050 A1B scenario) / 20c3m climatology
Downscale, and remove bias
- Interpolate temperature and precipitation anomalies with a first-order bilinear spline technique across an extent larger than our high-resolution climatology dataset. We use a larger extent to account for the climatic variability outside of the bounds of our final downscaled extent.
- Our GCM anomaly datasets are now at the same spatial resolution as our high resolution climatology dataset.
For temperature, add the interpolated temperature anomalies to the high-resolution climatology data (e.g., PRISM). For precipitation, multiply the interpolated anomalies by the climatology data. This step effectively downscales the data and removes model biases by using observed data values as baseline climate.
- The final products are monthly downscaled high resolution (2-km or 771-m for PRISM) data.

Why a proportional anomaly for precipitation?

A proportional anomaly for precipitation reduces the possibility of projecting negative precipitation in the future. Negative projections could occur with absolute anomalies if wet biases in the model lead to circumstances where the absolute reduction in precipitation is greater than the observed average precipitation.

Complications can arise with the proportional method in arid areas (where climatological precipitation is close to zero) when dividing by a very small number could produce unreasonably large precipitation increases. In the rare events that this does occur, we truncate the top 0.5% of anomaly values to the 99.5 percentile value for each anomaly grid.

This results in:

no change for the bottom 99.5% of values,
little change for the top 0.5% in grids where the top 0.5% of values are not extreme, and
substantial change only when actually needed, such as cases where a grid contains one or more cells with unreasonably large values resulting from dividing by near-zero.

We don’t omit precipitation anomaly values of a certain magnitude. Instead, we use a data distribution-based quantile to truncate the most extreme values. The 99.5% cutoff was chosen after consideration of the ability of various quantiles to capture extreme outliers. This adjustment allows the truncation value differ for each grid, because it is based on the distribution of values across a given grid.

References

Climatic Research Unit (CRU), University of East Anglia, accessed 16 July 2014.

Hay LE. 2000. A comparison of delta change and downscaled GCM scenarios for three mountainous basins in the United States. Journal of the American Water Resources Association 36(2) 387-397.

Fowler HJ, Blenkinsop S, Tebaldi C. 2007. Linking climate change modelling to impacts studies: recent advances in downscaling techniques for hydrological modelling. International Journal of Climatology 27:1547-1578.

Hayhoe KA. 2010. A standardized framework for evaluating the skill of regional climate downscaling techniques. PhD thesis, University of Illinois. Accessed 2 July 2014.

McAfee SA, Walsh JE, Rupp TS. 2013. Statistically downscaled projections of snow/rain partitioning for Alaska. Hydrological Processes.

McAfee SA, Russell JL, Webb RS. 2012. Influence of bias correction on simulated landcover changes. Geophysical Research Letters 39:L16702.

PRISM Climate Group, Oregon State University, accessed 4 Sept 2011.

Prudhomme C, Reynard N, Crooks S. 2002. Downscaling of global climate models for flood frequency analysis: where are we now? Hydrological Processes 16:1137-1150.

Walsh JE. 2011. Video: Global climate model performance in the Arctic: opportunities and challenges for downscaling. Presented at the Climate Data Downscaling workshop held in Anchorage, AK. Produced in collaboration with ACCAP.

Walsh JE, Chapman WL, Romanovsky V et al. 2008. Global Climate Model Performance over Alaska and Greenland. Journal of Climate 21 (23): 6156–74.