"ClimDown" is a Climate Downscaling package for the R statistical programming language. It was written at the Pacific Climate Impacts Consortium (PCIC) with support from Environment and Climate Change Canada.
The package provides routines for statistical downscaling of coarse scale global climate model (GCM) output to a fine spatial resolution.
PCIC's suite of routines include several different (yet related) downscaling techniques. The entire process is named Bias Correction/Constructed Analogues with Quantile mapping reordering (BCCAQ) and is composed of the following steps.
- Constructed Analogues (CA)
- Climate Imprint (CI)
- Quantile Delta Mapping (QDM)
- Rerank
See refer to the package documentation for details on each step and references to the corresponding scientific literature.
Changes in global climate have widespread impacts on the environment, economic activity, and human health, especially in high latitudes where warming is proceeding more rapidly and where ecosystems and traditional lifestyles are particularly sensitive to the impacts of warming.
Planning for adapting to climate change requires scientifically sound information about the future climate. Global climate models (GCMs) simulate future climate under different emission scenarios. However, GCMs simulate average conditions over large grid cells--typically on the order of 10,000 square kilometers or more per cell--which is often too coarse a resolution for regional and local applications. The use of original GCM data is not always the best option to provide adaptation-relevant information at the local scale.
Bias in model simulated local climate is of concern for many applications. For example, compared with observations, the median temperature simulated by GCMs from the Coupled Model Intercomparison Project Phase 5 (CMIP5) shows biases relative to Climate Research Unit high-resolution gridded dataset (CRU TS3.10) ranging from -3° C to 1.5° C for seasonal and annual mean temperatures in 26 global land areas (Flato et al. 2013). Precipitation simulated by CMIP5 models is also biased relative to observations (Flato et al. 2013). These biases hinder the direct application of model simulated future climate for impacts modelling and adaptation planning since climate impacts are often related to certain physical or biophysical thresholds. As a result, adaptation planning often uses model simulated future climate information that has incorporated some sort of downscaling and bias correction. Additionally, climate information is more useable and is less prone to misinterpretation when presented in a manner specific to the local community and/or impacts most relevant to a particular sector. High-resolution future projections of impact-relevant climate indices can be particularly useful in this regard.
ClimDown has been used to produce such bias corrected, downscaled GCMs for current and future climates, and could be used to do so for anywhere else in the world.
You can install the latest ClimDown release from CRAN using the R
interpreter: > install.packages('climdex.pcic')
If you are interested in a development version or a specific release of ClimDown, you can use the devtools package as an installation alternative.:
> install.packages('devtools')
> devtools::install_github("pacificclimate/ClimDown", ref="release")
# Or
> devtools::install_github("pacificclimate/ClimDown", ref="1.0.1")
ClimDown reads all of its input and produces its output in NetCDF format and manages numeric units using the UDUNITS2 library. The NetCDF and udunits2 libraries are system dependency of the package. Ensure that the following packages are installed under Debian/Ubuntu Linux systems: libnetcdf-dev, netcdf-bin, and libudunits2-dev. For other systems, follow the installation NetCDF install instructions and UDUNITS2 install instructions provided by Unidata.
The BCCAQ algorithm implemented by ClimDown is a complex, multi-stage operation, and as such performance will vary widely depending on the size of the input, the degree of parallelism selected by the user, and the performance characteristics of user's system (CPU speed, available RAM, I/O speed).
We have previously written about some of the complexities involved in downscaling performance, but it remains an area of active study.
Consider our experience as a matter of anecdote. We typically run ClimDown for downscaling 150 year, daily GCM simulations to a Canada-wide ANUSPLIN grid (approximately 10km resolution, 1068 by 510 cells). On our Linux systems, such runs can take up to 7 days to complete. However each of the different downscaling steps has different opportunities for parallelism and different performance characteristics. Typical of our runs is something like this:
- CI: 1 core, 10 GB RAM, Run time ~ 7 hours
- CA: 8 cores, 10 GB RAM, Run time ~ 2 hours
- QDM: 1 core, 36 GB RAM, Run time ~ 1.5 days
- rerank: 4 cores, 8 GB RAM, Run time ~ 4 days
In general, this downscaling technique is very expensive for large spatiotemporal domains. The more you can limit your domain, the faster your runtime will be. For small domains, it may be possible to run ClimDown on a typical workstation, but in general we do all of our production runs on rack-mounted supercomputers.
Though Windows binaries for ClimDown are available from CRAN, no effort has been made to optimize this package for Windows and your mileage may vary.