High-resolution and bias-corrected CMIP5 projections for climate change impact assessments

www.nature.com/scientificdata High-resolution and bias-corrected Data Descriptor CMIP5 projections for climate change impact assessments OPEN Carlos Navarro-Racines1,2, Jaime Tarapues1,2, Philip Thornton2,3, Andy Jarvis1,2 & Julian Ramirez-Villegas 1,2,4* Projections of climate change are available at coarse scales (70–400 km). But agricultural and species models typically require finer scale climate data to model climate change impacts. Here, we present a global database of future climates developed by applying the delta method –a method for climate model bias correction. We performed a technical evaluation of the bias-correction method using a ‘perfect sibling’ framework and show that it reduces climate model bias by 50–70%. The data include monthly maximum and minimum temperatures and monthly total precipitation, and a set of bioclimatic indices, and can be used for assessing impacts of climate change on agriculture and biodiversity. The data are publicly available in the World Data Center for Climate (WDCC; cera-www. dkrz.de), as well as in the CCAFS-Climate data portal (http://ccafs-climate.org). The database has been used up to date in more than 350 studies of ecosystem and agricultural impact assessment. Background & Summary There is a variety of methods to project the impacts of climate change on agriculture and biodiversity. This diversity arises, at least in part, from the difficulty to couple local-scale agricultural or species distribution and abundance models with General Circulation Model (GCM) projections, which are inherently uncertain1–3. GCMs can only model earth processes in coarse grid-cells, which are unsuitable for local agricultural studies4,5. Most impact models for agriculture and biodiversity require high-resolution environmental data6,7. Some authors (e.g. refs. 8–10) argue that original GCM resolutions should be kept so as not to bias or alter the physical plausibility of GCMs. Nevertheless, agricultural and natural landscapes have large spatial variations, particularly in the tropics, where orography, climate (especially precipitation), soils and crop management, vary across small distances11. The vast majority of agricultural and biodiversity researchers have used downscaling in impact studies6,12 (but see refs. 13,14). This is because conservation plans, niche models, crop models, and biodiversity evaluation require high resolution inputs. Downscaling and bias correction of climate model output produces data that allows local rather than regional or global projections of climate change and its impacts15,16. Planning, modeling and monitoring can therefore be at municipality, watershed or other sub-national scales17–21. Downscaling techniques range from smoothing and interpolation of GCM anomalies19, to statistical modeling, neural networks, and regional dynamical climate modelling22. They differ in accuracy, output resolution, computational requirements and climatic science robustness. Dynamical and statistical downscaling are the most frequently used techniques to downscale GCMs for agricultural impact studies23,24. Bias-correction, on the other hand, focuses on using different types of statistical techniques to make the climate model output more realistic, and, in many cases (i.e. when observations are available at high spatial resolution), also of greater spatial resolution15,25. Dynamical downscaling uses Regional Climate Models (RCMs) to increase the resolution of climate projections, with boundary and initial conditions from a GCM as inputs26–28. RCMs consider more detailed specifications of land use and water bodies, simulate mesoscale processes in more detail than GCMs, and, in some cases, are capable of explicitly resolving convective rainfall processes29,30. RCMs are computationally expensive, and require physical understanding of the climate system, time and storage to obtain a single scenario-by-period output21. RCM outputs have been made available recently through the Coordinated Regional Climate Downscaling Experiment (CORDEX)31. However, given their computational cost, only a handful of RCM–GCM combinations 1 International Center for Tropical Agriculture (CIAT), Cali, Colombia. 2CGIAR Research Program on Climate Change, Agriculture and Food Security (CCAFS), c/o CIAT, Cali, Colombia. 3International Livestock Research Institute (ILRI), Nairobi, Kenya. 4School of Earth and Environment, University of Leeds, Leeds, UK. *email: Scientific Data | (2020) 7:7 | https://doi.org/10.1038/s41597-019-0343-8 1 www.nature.com/scientificdata/ www.nature.com/scientificdata can realistically be used to produce future high-resolution climate change projections32,33. Moreover, RCM outputs are also subject to climate model error from both the structure of the RCM and the boundary conditions of the driving GCM30,34. Statistical downscaling (SD) is an easier and computationally less expensive method to develop climate change projections with high spatial resolution35. SD typically consists of two steps, (i) developing a statistical relationship between local climate variables and large-scale predictors, and (ii) the application of those statistical models onto future GCM output to derive future downscaled data36. SD assumes that climates will only change at coarse scales and that relationships between variables at local scale remain relatively constant in the future period30,36. Bias correction (BC) is yet simpler than SD, and is typically implemented by applying a ‘change factor’ or ‘delta’ derived from a GCM onto the historical observations15,35. BC can also be implemented by applying a ‘nudging’ factor to the climate model output, or by quantile-mapping of climate model outputs onto observations16,37. Since no GCM is a perfect representation of the true climate, BC seeks to correct those attributes in the climate model output that are known or hypothesized to be important for impacts modeling4,15,38. Here, we used BC to develop the CCAFS-Climate global database of bias corrected climate change projections. To develop the database, we applied the delta method (a simple BC) to 35 Coupled Model Intercomparison Project Phase 5 (CMIP5) models39, and four representative concentrations pathways (RCPs)40. We used the delta method since we focus on providing data for 30-year mean climate conditions, and because the method has already been shown to be robust to correct mean climate conditions in other regions15. For each GCM, we used the 30-year future periods named as 2030s (2020–2049), 2050s (2040–2069), 2070s (2060–2089) and 2080s (2070–2099) and three climate variables (mean monthly maximum and minimum temperatures and monthly rainfall). We used the WorldClim global climate database11 as the reference set of observations for the historical period. The database is freely available through the World Data Center for Climate (WDCC; cera-www.dkrz.de)41, as well as through the CCAFS-Climate data portal (http://ccafs-climate.org). We evaluate t (...truncated)