Drylands are home to more than 38% of the total global population and are one of the most sensitive areas to climate change and human activities1, 2. Projecting the areal change in drylands is essential for taking early action to prevent the aggravation of global desertification3, 4. However, dryland expansion has been underestimated in the Fifth Coupled Model Intercomparison Project (CMIP5) simulations5 considering the past 58 years (1948–2005). Here, using historical data to bias-correct CMIP5 projections, we show an increase in dryland expansion rate resulting in the drylands covering half of the global land surface by the end of this century. Dryland area, projected under representative concentration pathways (RCPs) RCP8.5 and RCP4.5, will increase by 23% and 11%, respectively, relative to 1961–1990 baseline, equalling 56% and 50%, respectively, of total land surface. Such an expansion of drylands would lead to reduced carbon sequestration and enhanced regional warming6, 7, resulting in warming trends over the present drylands that are double those over humid regions. The increasing aridity, enhanced warming and rapidly growing human population will exacerbate the risk of land degradation and desertification in the near future in the drylands of developing countries, where 78% of dryland expansion and 50% of the population growth will occur under RCP8.5.
Drylands are defined as regions where precipitation is counterbalanced by evaporation from surfaces and transpiration by plants (evapotranspiration)3. Because most dryland soil is relatively infertile and the vegetation cover is sparse, dryland ecosystems are substantially more fragile1. Desertification and degradation are pervasive in drylands owing to global warming and the effects of rapid economic development, population growth and urbanization8. There are also some studies indicating that the increasing hydroclimatic intensity will become a predominant signature of twenty-first-century warming, which leads to shorter, less frequent, and less widespread precipitation events and an increase in the length of dry spells9. These trends may induce the expansion of drylands and further increase the fraction of the population that is affected by water scarcity and land degradation1, 4. Knowledge of how climate change will affect the extent of drylands in the future is essential for their protection and for adaptation strategies10. The CMIP5 has generated projections using several emissions scenarios11 and has provided a crucial reference for maintaining drylands as renewable resources. This study verifies CMIP5 simulations and bias-corrects the projections using historical observational data to provide a clear understanding of the spatial and temporal evolution of drylands in the future. The results may motivate decision makers to respond early and effectively to mitigate the pending global desertification.
The aridity of a region is generally measured by the aridity index (AI), which is the ratio of total annual precipitation to potential evapotranspiration (PET). Under this quantitative indicator, drylands are defined as regions with AI < 0.65 and are further divided into subtypes of hyper-arid (AI < 0.05), arid (0.05 ≤ AI < 0.2), semiarid (0.2 ≤ AI < 0.5) and dry subhumid (0.5 ≤ AI < 0.65) regions3. The observational data used here are from the Climate Prediction Center (CPC; refs 12,13). The simulation data are from 20 global climate models of CMIP5 (ref. 11; Methods). As the ensemble mean of these CMIP5 models (CMIP5-EM) can filter the uncertainty from inter-model variability and is the best representation of the response to imposed external forcing, it is better at predictions than any individual member14, 15 and is used to reflect the simulated aridity changes in this study.
To ensure the reliability of the future projections (2006–2100), it is critical to evaluate CMIP5-EM historical simulations (1948–2005) of dryland variability compared with observations over the same time period5. The historical values of the global observed AI and CMIP5-EM values over 58 years are compared in Table 1 and Fig. 1. The observed AI decreased remarkably, with a mean net trend of −0.050 per 58yr; the areas with drying trend cover up to 66% of the global land area (Table 1). By contrast, the mean trend of CMIP5-EM is only −0.012 per 58yr, which is approximately one-fourth of the observed trend; drying regions cover only 59% of the global land area. A subset of the AI data over the last 15 years of the historical period (1991–2005) is compared with the first 15 years (1948–1962) to highlight the temporal changes (Table 1). The observed areal increases in hyper-arid, arid, semiarid and subhumid land types from neighbouring wetter subtypes are 0.62%, 1.16%, 2.32% and 3.32%, respectively, of the global land area, whereas the increases according to CMIP5-EM are 0.05%, 0.14%, 0.37% and 0.50%, respectively. Similarly, the decreases in the subtype areas from drier to neighbouring wetter subtypes in CMIP5-EM are approximately one-third of those of the observations.
Table 1: Comparison of the AI for the observed data and CMIP5-EM simulation.
The PRECipitation REConstruction over Land (PREC/L) data set developed by the CPC at a spatial resolution of 0.5° is involved in this study, which is interpolated from observations of the Global Historical Climatology Network (GHCN) version 2 and the Climate Anomaly Monitoring System (CAMS) data set for the period extending from 1948 to present12. The construction of the data set is introduced in Supplementary Section 1. In addition, the PET data set is provided by Feng and Fu5, which is calculated using the Penman–Monteith method31, 32. The surface air temperature (SAT) data set used to calculate PET is also from CPC, which is labelled as the GHCN_CAMS Gridded 2 m temperature13. The solar radiation, specific humidity and wind speed reanalysis data sets used are from the Global Land Data Assimilation System (GLDAS; ref. 33). The algorithms used for PET and the uncertainty and reliability of the Penman–Monteith method are discussed in Supplementary Section 2.
The precipitation and PET simulation data set used are provided by Feng and Fu5. It is derived from monthly mean temperature, precipitation, solar radiation, specific humidity and wind speed products from 20 CMIP5 climate models11 (Supplementary Table 2). Most of these simulations cover the period from 1850 to 2005 (ref. 11). Here, we analyse only the period from 1948 to 2005, for which the CMIP5, CPC and GLDAS data sets are commonly available. Because the CMIP5 models have different spatial resolutions, the simulated fields are statistically downscaled to a 0.5° × 0.5° resolution to match the observational data sets5. Moreover, to focus on the temporal variation and long-term climate change, the model simulations are adjusted to have the same climatology of 1961–1990 as the observations5.
GPP data.
The gross primary productivity (GPP) is the total amount of carbon fixed in the process of photosynthesis by plants in an ecosystem. A complete 11-year period (2000–2010) of yearly GPP data were acquired from NASA Goddard Space Flight Center (http://ladsweb.nascom.nasa.gov/data/search.html) to analyse the relationship between the AI and GPP. The yearly GPP data (MOD17A3) are retrieved at a spatial resolution of 1km × 1km as a part of the Terra satellite’s Moderate Resolution Imaging Spectroradiometer (MODIS) level-4 collection 5.5 (C055), which was recently updated with yearly gridded land products. The data are processed by the MODIS Reprojection Tool (MRT) with a spatial grid cell resolution of 0.5° × 0.5°.
Population data.
The population counts are from the Gridded Population of the World, version 3 (GPWv3) (http://sedac.ciesin.columbia.edu/gpw/index.jsp). A proportional allocation gridding algorithm, utilizing more than 300,000 national and sub-national administrative units, is used to assign population values for 2000 to grid cells with a spatial resolution of 0.5° × 0.5°. The population counts for 2025 are generated using the Country-level Population and Downscaled Projections Based on the Special Report on Emissions Scenarios (SRES) B2 Scenario (1990–2100) data set; CIESIN’s Gridded Population of World, version 2 (GPWv2), is used as the base map34.
Because the CMIP5 and observed data over the historical period both have too many degrees of freedom, establishing a mapping relationship between them in such a high-dimensional space is impractical. By applying principal component analysis (PCA), the CMIP5 and observed data sets are both reduced to a lower dimensional space and decomposed into eigenvectors and corresponding principal components (PCs) (Supplementary Fig. 1). Each PC of the CPC data set is regressed by the leading PCs of CMIP5, and a matrix of regression coefficients is obtained. According to this relationship, the observed PCs can be predicted by the simulated PCs of CMIP5 in the future. The leading PCs are considered as correction factors not only because they are responsible for most of the variance in mathematics, but they also present high correlation with the patterns induced by external forces and by internal climate variability35. Although the correlation does not imply causality, it may reveal the possible physical meanings underlying each PC (see discussion in Supplementary Section 4.1). This method has been used to correct the seasonal predictions18, 19, 20, and we adapted it in this study to correct the long-term changes of the CMIP5 projections.
The schematic of correction is shown in Supplementary Fig. 2 and the practical procedures are introduced as follows:
First, the simulated fields are corrected by subtracting the climate drift, which is estimated as the difference between the forecasts and the observations averaged over climatic timescales. Then the simulated field Zsim(t) and observed field Zobs(t) can both be expressed as a matrix of size m × n, where m is the number of spatial grid points and n is the length of the time series. Generally, m is much larger than n. To reduce the spatial dimensions, we expand Zsim(t) and Zobs(t) by PCA:
where Xi and Yj are spatial patterns with a dimension of m; bi and aj are PCs corresponding to each Xi and Yj with a dimension n. Using multiple linear regression, each aj(t) can be regressed by the K leading bi(t) (i = 1,2, …, K):
where ci, j is the regression coefficient and ej is the regression residual. Projecting Zsim(n + 1) (the predicted field at the time of n + 1) onto each spatial pattern Xi will provide the corresponding time coefficient bi(n + 1):
By inserting bi(n + 1) into equation (3), we can obtain a prediction of the observed time coefficient
Then, the adjusted prediction of Zobs at a time of n + 1 can be obtained by combining the spatial pattern:
Generally, the climate models have a poor ability to simulate high-frequency variability, and retaining higher-order PCs may introduce additional noise. In this case, the optimal number of leading PCs should be examined. Here the widely used leave-one-out cross-validation during the historical period is conducted to choose the cumulative number of leading PCs (see detailed procedures in Supplementary Section 4.2 and Supplementary Fig. 3).
Because the corrected AI contains information from both observations and CMIP5, it is considered to be more accurate than the original projections. The validity of this method is verified by a posteriori independent validation (see details discussion in Supplementary Section 5), and the results confirm that the correction method is robust and reliable. On this basis, the PCA adjusting method is applied to the future period in the same way. The 58-year record of observations and simulations during the historical period is decomposed by PCA, and the linear regression is obtained. The simulation field for each year during the future period is projected onto the decomposed spatial patterns to obtain a time coefficient. Applying this time coefficient to the regression relationship will provide a ‘predicted’ time coefficient for the observations and allow a ‘prediction’ of the observed field. The adjustment is conducted for each year from 2006 to 2100 by obtaining the adjusted predictions under RCP8.5 and RCP4.5.
Graphics software.
All maps and plots were produced using licensed IDL version 8.2.
Reynolds, J. F.et al. Global desertification: Building a science for dryland development. Science316, 847–851 (2007).
