Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Selenium deficiency risk predicted to increase under future climate change

Selenium deficiency risk predicted to increase under future climate change Selenium deficiency risk predicted to increase under future climate change a a b c a,d e b Gerrad D. Jones , Boris Droz , Peter Greve , Pia Gottschalk , Deyan Poffet , Steve P. McGrath , Sonia I. Seneviratne , f a,d,1 Pete Smith , and Lenny H. E. Winkel a b Eawag, Swiss Federal Institute of Aquatic Science and Technology, 8600 Duebendorf, Switzerland; Institute for Atmospheric and Climate Science, ETH c d Zurich, 8092 Zurich, Switzerland; Potsdam Institute for Climate Impact Research, 14473 Potsdam, Germany; Institute of Biogeochemistry and Pollutant Dynamics, ETH Zurich, 8092 Zurich, Switzerland; Department of Sustainable Soils and Grassland Systems, Rothamsted Research, Harpenden AL5 2JQ, United Kingdom; and Institute of Biological and Environmental Sciences, School of Biological Sciences, University of Aberdeen, Aberdeen AB24 3UU, United Kingdom Edited by Jerome Nriagu, University of Michigan, Ann Arbor, MI, and accepted by Editorial Board Member David W. Schindler January 6, 2017 (received for review July 15, 2016) Deficiencies of micronutrients, including essential trace elements, were (i) to test hypothesized drivers of soil Se concentrations, (ii)to affect up to 3 billion people worldwide. The dietary availability of predict global soil Se concentrations quantitatively, and (iii)to trace elements is determined largely by their soil concentrations. quantify potential changes in soil Se concentrations resulting from Until now, the mechanisms governing soil concentrations have been climate change. To achieve these objectives, several regional- to evaluated in small-scale studies, which identify soil physicochemical continental-scale datasets reporting total soil Se concentrations [n = properties as governing variables. However, global concentrations of 33,241 data points (5, 14–29); see SI Materials and Methods for trace elements and the factors controlling their distributions are dataset details] and 26 environmental variables describing climate, virtually unknown. We used 33,241 soil data points to model recent soil physicochemical properties, irrigation, water stress, erosion, (1980–1999) global distributions of Selenium (Se), an essential trace runoff, land use, soil type, lithology, bedrock depth, vegetation/ element that is required for humans. Worldwide, up to one in seven canopy characteristics, and population density (30–42) (see Table people have been estimated to have low dietary Se intake. Contrary S1 for details of variables) were assessed to make global predictions to small-scale studies, soil Se concentrations were dominated by of soil Se concentrations for recent (1980–1999) and future (2080– climate–soil interactions. Using moderate climate-change scenarios 2099) periods. Predictions were made using three machine-learning for 2080–2099, we predicted that changes in climate and soil organic tools: one randomForest (RF) model and two artificial neural carbon content will lead to overall decreased soil Se concentrations, network models, herein referred to as “predictive models.” Addi- particularly in agricultural areas; these decreases could increase the tionally, structural equation modeling (SEM) was used to evaluate prevalence of Se deficiency. The importance of climate–soil interac- potential mechanisms independently and to quantify complex in- tions to Se distributions suggests that other trace elements with sim- teractions between Se and the relevant predictor variables. ilar retention mechanisms will be similarly affected by climate change. Results and Discussion selenium soils global distribution prediction climate change | | | | After variable selection, seven variables were retained and were considered the most important factors controlling soil Se con- icronutrients are essential for maintaining human health, centrations: the aridity index (AI, unitless) [i.e., the ratio of po- Mand although they are needed in only trace amounts, de- tential evapotranspiration (PET, mm/d) to precipitation (mm/d)]; ficiencies reportedly affect 3 billion people worldwide (1, 2). One such micronutrient is selenium. Inadequate dietary Se intake af- Significance fects up to 1 in 7 people and is also known to affect livestock health adversely (3, 4). Because dietary Se intake depends largely on Se The trace element selenium is essential for human health and is content in soil and bioavailability to crops (5–7), understanding the required in a narrow dietary concentration range. Insufficient mechanisms driving soil concentrations and predicting global dis- selenium intake has been estimated to affect up to 1 billion tributions could help prevent Se deficiency (8). However, global people worldwide. Dietary selenium availability is controlled by soil Se concentrations and the factors affecting Se distributions are soil–plant interactions, but the mechanisms governing its broad- largely unknown (9). Apart from soils, Se is present in all other scale soil distributions are largely unknown. Using data-mining environmental compartments [i.e., the lithosphere, hydrosphere, techniques, we modeled recent (1980–1999) distributions and biosphere, and atmosphere (9)], which all play a role in global Se identified climate–soil interactions as main controlling factors. biogeochemical cycling and distribution (7, 10). Furthermore, using moderate climate change projections, we The factors driving soil Se concentrations [e.g., increased sorption predicted future (2080–2099) soil selenium losses from 58% of with decreased pH and soil reduction potential (Eh) and increased modeled areas (mean loss = 8.4%). Predicted losses from crop- clay and soil organic carbon (SOC) content; see Table S1 and refs. 7 lands were even higher, with 66% of croplands predicted to lose and 11 for a review of the previously reported drivers of soil Se] 8.7% selenium. These losses could increase the worldwide have been evaluated primarily through small-scale experimentation prevalence of selenium deficiency. (e.g., soil columns; see ref. 12); however, broad-scale distributions cannot be inferred from such studies. For example, soils in small- Author contributions: G.D.J., B.D., and L.H.E.W. designed research; G.D.J., B.D., D.P., and L.H.E.W. scale experiments are often manipulated [e.g., by carbon amend- performed research; G.D.J., B.D., P. Greve, P. Gottschalk, S.P.M., S.I.S., P.S., and L.H.E.W. ments (12)] to achieve desired conditions, obscuring the natural contributed data/soil samples; G.D.J., B.D., D.P., and L.H.E.W. analyzed data; and G.D.J., B.D., D.P., and L.H.E.W. wrote the paper with technical input from all authors. processes that may influence Se retention capacity. Additionally, climate variables, which likely affect soil Se concentrations directly The authors declare no conflict of interest. as a source (e.g., deposition; see refs. 8 and 13) or indirectly by This article is a PNAS Direct Submission. J.N. is a Guest Editor invited by the Editorial Board. affecting soil retention of Se (e.g., sorption), are ignored in small- scale experiments. Therefore, to predict the global distributions, Freely available online through the PNAS open access option. broad-scale analyses of soil Se drivers are essential. To whom correspondence should be addressed. Email: lwinkel@ethz.ch. Here we report on the influence of soil and climate variables on This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. worldwide Se distributions in soils 0–30 cm deep. Our objectives 1073/pnas.1611576114/-/DCSupplemental. 2848–2853 | PNAS | March 14, 2017 | vol. 114 | no. 11 www.pnas.org/cgi/doi/10.1073/pnas.1611576114 world. In sensitivity analyses, soil Se increased with increases in clay content and with decreases in soil pH (Fig. 2 and Figs. S3 and S4), both of which are known to increase soil Se sorption (7, 44). Al- though soil Se is known to partition/complex with organic matter (7), soil Se was affected only weakly by changes in SOC (Figs. S1, S3,and S4 and Table S2). Furthermore, changes in lithological classes resulted in negligible changes in soil Se when other variables were held constant (Fig. S4). Although lithology was of minor im- portance in this study, we recognize that it can influence soil Se concentrations at local scales (45). Climate Effects on Soil Se. Climate variables (i.e., AI, precipitation, and ET) were dominant factors driving soil Se concentrations, likely because they control leaching from soils, and observed pat- terns within the sensitivity and SEM analyses for all climate vari- ables are consistent with this hypothesis. High precipitation and AI negatively affected soil Se concentrations, whereas ET positively affected soil Se concentrations (Fig. 2, Figs. S3 and S4,and Table S2). Although AI (i.e., the ratio of PET to precipitation) and precipitation are inversely related, both variables exerted negative effects on soil Se, suggesting that separate mechanisms drive these patterns. Although precipitation increases the transport of dis- solved Se species in soil solution by increasing vadose zone flow, AI likely affects leaching by controlling soil redox conditions and thus Se speciation, sorption, and mobility. It has been reported that as AI increases (i.e., PET increases relative to precipitation), soils become drier (46), resulting in more oxidizing soil conditions (47). Oxidized Fig. 1. Summary of the processes governing soil Se concentrations. Domi- Se species (e.g., oxyanions) are more soluble and mobile than re- nant processes (and bulleted examples) governing soil Se concentrations are duced species (e.g., selenides) (7, 11, 12, 48). Therefore, soil drying indicated. Text colored in red, green, and blue indicates processes affecting soil Se losses, retention, and sources/supplies, respectively. Factors re- likely increases the presence of oxidized/mobile soil Se species, sponsible for increases (+) and/or decreases (−) in soil Se as well as processes which can be leached during subsequent rain events. not explicitly examined in our analysis (*) are indicated. Soil drying likely increases Se mobility but also can reduce soil Se transport (although these processes likely occur at different time scales). Leaching is driven by the ratio of ET to precipitation clay content (%); evapotranspiration (ET, mm/d); lithology; pH; (also known as the “evaporative index,” EI) (49). As previously precipitation (mm/d); and SOC (0–30 cm depth, tons of C/ha). mentioned, precipitation increases the transport of Se through the With these variables, the accuracy of the predictive model was vadose zone, but as ET increases relative to precipitation, more 2 2 high (average R = 0.67, average cross-validation R = 0.49, n = moisture is removed from the soil column. This removal of moisture 1,000 iterations for each model) (Fig. S1), and the precision was high [based on a low SD of the modeled prediction (0.032 mg Se/kg) relative to the mean (0.35 mg Se/kg)] (Fig. S2). For the SEM 1.2 1.2 analysis, both the standardized root mean squared residual A B 1.0 (SRMR) (0.043) and the comparative fit index (CFI) (0.962) in- Zone 1 1.0 Zone 2 dicated a good fit between the observed and modeled data (43). 0.8 0.8 Zone 3 All variables retained within the SEM analysis were statistically 0.6 0.6 significant (i.e., P < 0.05) (Fig. S3), and, based on the predictive 0.4 0.4 model sensitivity analyses and the SEM regression weights, the 0.2 0.2 modeling results corroborated each other strongly (see following 0.0 0.0 section), suggesting that processes driving soil Se concentrations 048 12 468 10 Aridity Index (AI) [unitless] pH were described accurately. Because the data were averaged on a 1° 1.2 1.2 scale, this modeling approach likely captures the broad-scale C 4.2 D 0.5 pH Evapotranspiraon 4.8 1.0 1.0 1.0 (ET) [mm/d] mechanisms but potentially misses important local-scale factors. 5.3 1.5 5.9 2.0 0.8 0.8 Soil Se concentrations were determined largely by interactions 6.4 2.5 ET = precipitaon between climate and soil variables (Figs. 1 and 2, Figs. S1, S3,and 7.0 3.0 0.6 0.6 7.5 3.5 S4,and Table S2). In the SEM, AI and precipitation had the 4.0 8.1 0.4 0.4 4.5 8.6 greatest direct and indirect effect, respectively, on soil Se concen- 9.2 5.0 0.2 0.2 trations (Table S2). Basedonaveragedrelative importancefrom 0.0 0.0 the predictive models (Fig. S1), AI was the most important pre- 0.1 1 10 0123456 dictor (100 ± 0.3%) followed by pH (60 ± 0.7%), precipitation (58 ± Aridity Index (AI) [unitless] Precipitaon [mm/d] 1%), ET (50 ± 0.8%), clay content (45 ± 0.4%), lithology (33 ± Fig. 2. Univariate and bivariate sensitivity analyses of the predictive models. 0.5%), and SOC (29 ± 0.5%). Sensitivity analyses were performed (A and B) The independent effects of AI (A)and precipitation (B)weremod- on these variables to determine if the mechanisms driving soil Se eled by holding all other variables constant at the zonal averages as defined by concentrations changed in different zones represented by different thetwo-stepclustering. (C and D) Similarly, bivariate interactions between AI environmental conditions (see SI Materials and Methods for a de- and clay (C) and between precipitation and ET (D) are illustrated. These pa- scription). The soil Se patterns were similar between different zones, rameters were allowed to vary between the minimum and maximum observed suggesting that Se drivers were consistent regardless of the envi- value while all other variables were held constant at the mean value of the ronment (Fig. 2 and Fig. S4). This result suggests that the models entire dataset (n = 1,642). The dotted line in D indicates the conditions in can be used to predict soil Se concentrations in other regions of the which ET = precipitation. Other bivariate interactions are presented in Fig. S3. Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2849 Modeled soil selenium [mg/kg] ENVIRONMENTAL SCIENCES reduces the vadose zone flow, which in turn reduces Se mass sensitivity and SEM analyses suggest that the direct (i.e., non- transport. In theory, when ET and precipitation are equal, leaching mediated) effect of precipitation drives Se leaching from soils (Fig. 2 should be negligible. Although ET clearly dampens the negative and Figs. S3 and S4), thus explaining why low values resulted in high effects of precipitation in bivariate sensitivity analyses, a negative soil Se, even though the net effect is positive (Table S2). In bivariate relationship existed between precipitation and modeled Se despite sensitivity analyses, soil Se concentrations exceeded these values an ET:precipitation ratio of 1 (Fig. 2). This trend potentially could when low AI was modeled with low pH (1.12 mg Se/kg), when low be explained by plant Se uptake, which has been reported to precipitation was modeled with high clay content (0.86 mg Se/kg), increase with ET (6). Therefore, in addition to its positive indi- and when high clay content was modeled with low pH (0.56 mg Se/kg) rect effect by reducing leaching, ET also may have a direct negative (Fig. 2 and Fig. S3). Although Se concentrations were typically en- effect on soil Se by increasing plant uptake. This direct negative hanced in low-AI or low-precipitation environments, both variables effect was observed in the SEM analysis, but the relationship was could suppress the effects of other variables in high-AI or high- not statistically significant and therefore was removed. Although precipitation environments (Fig. 2 and Fig. S3). These results this negative effect may exist, it appears to be less important than demonstrate the dependence of soil Se concentrations on soil– the role ET plays in reducing Se leaching. Given the importance of climate interactions. Based on these analyses, low-Se soils are climate variables in governing soil Se concentrations, the observed most likely to occur in arid environments and in areas with high patterns between AI, precipitation, ET, and modeled soil Se con- pH and low clay content. Conversely, areas with low to moderate centrations strongly suggest that changes in climate will result in precipitation but relatively low aridity (e.g., cool and moist climates) changes in soil Se concentrations in time and space. and high clay content are likely to have higher soil Se concentrations. Climate–Soil Interactions and Soil Se. Although the direct effects of Predicted Global Soil Se Distributions. Global predictions were precipitation (i.e., leaching) were moderate, its indirect effects (i.e., made using models trained largely using temperate/midlatitude those mediated through other variables) were approximately three- datasets (Fig. S2). Although the available data adequately described fold larger (Table S2). Precipitation is known to affect soil forma- similar regions, data from tropical, extremely arid, and polar regions tion (i.e., pedogenesis), and in our analyses it strongly affected AI, were almost entirely absent (to the best of our knowledge, no broad- pH, ET, and clay content, which subsequently affect soil Se re- scale soil geochemical surveys are available from these regions). As tention (Fig. S3 and Table S2). Although there was a negative direct a result, predictions that were made for environments that fell effect between precipitation and soil Se, the sum of direct and in- outside our dataset’s domain were excluded (Fig. S5). Therefore, Se direct effects resulted in precipitation having a net positive effect predictions for 1980–1999 were retained for only 70% of land sur- (Table S2). Thus it is important to examine both direct and indirect faces. The majority of croplands and rangelands, which are areas of effects, because interpreting only total effects can lead to erroneous primary interest, given that soil Se concentrations and bioavailability conclusions about the mechanisms driving spatial patterns. in these regions largely drive the Se status in humans and livestock, In bivariate sensitivity analyses, both synergistic and antagonistic fall largely within the retained areas. interactions were observed and were strongest between aridity, Based on predictive models, the global mean soil Se concen- precipitation, clay content, and pH. In univariate sensitivity anal- tration for 1980–1999 was 0.322 ± 0.002 mg Se/kg (Fig. 3), similar yses, when all other variables were held constant, modeled soil Se to reported values (mean = 0.4 mg Se/kg; typical range 0.01–2mg concentrations were highest under low AI (0.83 mg Se/kg), low Se/kg) (50). Using this estimate, ∼13.1 million metric tons of Se precipitation (0.65 mg Se/kg), low pH (0.51 mg Se/kg), and relatively are stored in the top 30 cm of soil within the predicted area [i.e., 7 2 high clay content (0.47 mg Se/kg) (Fig. 2 and Fig. S4). It is important ∼70% of world’sland surface (1.04 × 10 km ); see SI Materials to note that, as long as PET is sufficiently low, environments with and Methods for the calculation]. Compared with other regions, low precipitation can have low AI values also. Furthermore, predicted soil Se concentrations were generally higher (typically Modeled Soil Se 1980-1999 <0.1 mg/kg 0.1-0.2 mg/kg 0.2-0.3 mg/kg 0.3-0.4 mg/kg 0.4-0.5 mg/kg >0.5 mg/kg Avg. = 0.32 mg/kg % change in Fig. 3. Geographical representation of the pre- soil Se 1980-1999 dictive modeling on a 1° scale. Maps illustrate the to 2080-2099 modeled soil Se concentrations (1980–1999) (A)and <-10% percentage change in soil Se concentrations be- -10--2.5% tween recent and future (2080–2099) conditions (B) -2.5-2.5% as a function of projected changes in climate (RCP6.0 2.5-10% scenario) and SOC content (ECHAM5-A1B scenario). >10% Predictions represent the average of the predictive Changes due to projected changes Avg. = -4.3% models and were based on the AI, soil clay content, in climate variables and SOC ET, lithology, pH, precipitation, and SOC. 2850 | www.pnas.org/cgi/doi/10.1073/pnas.1611576114 Jones et al. >0.2 mg Se/kg) (Fig. 3) in temperate and northern latitudes. In amount that is ∼20–30% of the total estimated Se mass that is cy- wet equatorial regions, concentrations were typically 0.3–0.5 mg Se/kg. cled yearly through the troposphere [i.e., 13,000–19,000 tons (as- Relatively low-Se soils (<0.2 mg Se/kg) were predicted for 15% of sumed to be metric tons)/y (10)], although our estimate is only for modeled areas and were restricted primarily to arid and semiarid 48% of the land surface. Our modeling approach is not a mass regions in Argentina, Australia, Chile, China, southern Africa, and the balance model, so Se fate could not be investigated. Nonetheless, southwestern United States. In some of these countries, low Se changes in Se concentrations in other environmental compartments content in crops and livestock has been reported (3), but it is im- are known from the past. For example, marine Se concentrations portant to note that many factors contribute to low Se content in throughout various periods of the Phanerozoic eon have been 1.5–2 plants (e.g., plant uptake pathways, soil Se speciation, and the orders of magnitude lower than current oceanic concentrations (53). 7 2 abundance of competing ions such as sulfate) (7). Based on areas with future predictions (7.19 × 10 km ), 58% of lands were predicted to lose soil Se (i.e., ΔSe less than −2.5%; Over- and Underpredictions of the Model. In an attempt to identify mean change = −8.4%); 20% were predicted to undergo minor potential missing variables, we examined the residuals of the changes (i.e., −2.5% < ΔSe < 2.5%; mean change = −0.3%); and predictive models. Spatial patterns of any missing variable should 22% were predicted to gain soil Se (i.e., ΔSe > 2.5%; mean change match those of the residuals (Fig. S2). Overall, the models = 5.7%)asa result of changesin climate and SOC(Fig. 3).Pre- underpredicted soil Se concentrations (average residual = −0.036 ± dicted soil Se losses were driven largely by changes in AI, whereas 0.009 mg Se/kg), suggesting that Se sources may be missing from soil Se gains were driven largely by changes in precipitation and the model. On a localized level, soil Se concentrations appear to SOC (Fig. S1). Compared with the total land surface, croplands be underpredicted in regions adjacent to regions of high marine were expected to lose more soil Se. Based on future predictions for productivity (e.g., western Alaska, western Ireland, western Norway, 6 2 croplands (7.55 × 10 km ), 66% of lands were predicted to lose western England, and Wales) (Fig. S2) (51). Marine environments soil Se (ΔSe less than −2.5%; mean change = −8.7%); 15% were are thought to increase soil Se concentrations via wet deposition predicted to undergo minor changes (–2.5% < ΔSe < 2.5%; mean (10, 13), and atmospheric deposition of Se thus could explain some change = −0.4%); and 19% were predicted to gain soil Se (ΔSe > of the model’s underprediction. However, global spatial data do 2.5%; mean change = 7.3%) (Fig. 3 and Fig. S6). Global pasture not exist for Se deposition and thus could not be analyzed. We lands also were predicted to lose soil Se, but to a lesser extent than included population density as a potential proxy for anthropogenic croplands, suggesting that Se deficiency in livestock could increase. emissions, but this was one of the least important variables in 7 2 Based on future predictions for pasture lands (2.55 × 10 km ), the variable selection procedure. We evaluated a wide variety of 61% of lands were predicted to lose soil Se (ΔSe less than −2.5%; qualitative factors [e.g., specific agricultural soil types (e.g., paddy mean change = −8.0%); 19% were predicted to undergo minor soils), specific sedimentary depositional environments (e.g., glacial changes (−2.5% < ΔSe < 2.5%; mean change = −0.4%); and 21% deposits), coal power plants, carbonaceous shale deposits, and were predictedto gainsoilSe(ΔSe > 2.5% mean change = 8.2%) others] that may affect soil Se distributions; however, we found no (Fig. 3 and Fig. S6). Areas with notable losses (i.e., ΔSe less than consistent discernable link between these variables and the broad- −10%) include croplands of Europe and India, pastures of scale distribution of the model residuals. China, Southern Africa, and southern South America, and the Despite the underprediction, overall patterns of modeled soil southwestern United States (Fig. 3 and Fig. S6). Areas of notable Se distribution match the actual distribution quite closely (Fig. S2), gain (ΔSe > 10%) are scattered across Australia, China, India, and 71% of predicted values were within ±0.05 mg Se/kg of the and Africa (Fig. 3 and Fig. S6). observed value, indicating that a majority of the predicted data were relatively accurate. Furthermore, the sensitivity analysis and Temporal Changes in Soil Se. Although our analysis indicates that SEM trends closely match hypothesized mechanisms governing soil future climate change will likely result in widespread changes in Se concentrations reported in the literature (Fig. 2 and Figs. S3 and soil Se, it does not indicate rates of change. To understand the S4). This finding suggests that the models are largely accurate and temporal changes in soil Se concentrations better, we analyzed for capture the dominant processes controlling broad-scale Se distri- soil Se and SOC in a subset of agricultural samples collected from butions. Nevertheless, future studies could include additional pre- the Broadbalk Experiment (Rothamsted, United Kingdom) be- dictor variables, especially those that are currently unavailable, to tween 1865 and 2010. Soil samples were taken from a control plot provide better estimates of broad-scale soil Se. Furthermore, to (unfertilized since 1843) and two “wilderness” plots (a maintained overcome some of this study’s limitations, predictions could be grassland and woodland), which were converted from the control made on more local/regional scales using higher resolution data. plot in 1882 (SI Materials and Methods,and Table S3). The ac- cumulation of Se in soil through time was statistically greater in Modeled Losses of Future Soil Se. The interactions between pre- the wilderness plots than in the control plot [one-way analysis of cipitation and other soil/climate variables strongly suggest that covariance (ANCOVA); year: F(1, 31) = 20.7, P < 0.01; plot F(2, climate changes could drive changes in soil Se concentrations. To 31) = 17.3, P < 0.01]. When controlling for SOC, however, there assess the influence of changes in climate and SOC, soil Se was were no statistical differences between the plots (ANCOVA; modeled for 2080–2099 using moderate climate change scenarios SOC: F(1, 30) = 10.7, P < 0.01; year: F(1, 30) = 6.0, P < 0.05; plot: [Representative Concentration Pathways (RCP) 6.0 for pre- F(2, 30) = 3.2, P > 0.05), indicating that increases in SOC were cipitation, AI, and ET (52) and European Centre/Hamburg driving soil Se accumulation in the model (Fig. S7), as is consistent Model (ECHAM) 5-A1B for SOC (33)]. Other climate scenarios with the results of the future modeling (Fig. S1). Natural changes in (e.g., RCP 8.5) were not used because SOC data were available soil Se concentrations previously have been hypothesized to occur only for A1B scenarios, which are most similar to RCP 6.0. over longer time scales (e.g., hundreds to thousands of years) (54); Future predictions were made for the entire globe, but, based on however, given that SOC and Se began to accumulate on these plots the filtering criteria used (SI Materials and Methods), predictions immediately after conversion, these results suggest that changes in were retained for ∼48% of theglobal land area. Basedonthese soil Se will follow environmental changes rapidly, perhaps on an pixels alone, soil Se concentrations were predicted to drop by 4.3% annual to decadal time scale. Between ∼1880 and 1980, soil Se on average, from 0.331 ± 0.003 mg Se/kg in 1980–1999 to 0.316 ± 0.002 mg Se/kg in 2080–2099, as a result of changes in climate and concentrations increased by ∼15, 35, and 60% on the control, SOC concentrations (Fig. 3). For soil at a depth of 0–30 cm, this loss grassland, and woodland plots, respectively (Fig. S7), indicating that corresponds to ∼403,763 tons of Se over 100 y, or 4,037.6 tons of Se the magnitude of changes predicted to occur by the end of the 21st lost peryear(see SI Materials and Methods for the calculation), an century is plausible. The rates of change in soil Se concentrations Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2851 ENVIRONMENTAL SCIENCES model was iterated 1,000 times using 90% of the data for model training and following environmental perturbations is largely unstudied and 10% of the data for cross-validation for each iteration. The training and cross- should be evaluated further to understand soil Se dynamics better. validation data were chosen at random for each iteration. The model predictions were averaged to estimate recent (1980–1999) global soil Se concentrations; Outlook however, predictions were considered valid only if the environmental parame- One of our aims was to identify the broad-scale mechanisms ters for each pixel fit within the domain of the observed data (Fig. S5). governing soil Se retention. Therefore, at a 1° resolution, the data Sensitivity analyses were performed during each iteration to investigate used are likely too coarse to evaluate or identify the influence of the independent effect of each variable on modeled soil Se concentrations. many small- to regional-scale factors (e.g., local sources, specific Based on all input variables, three environmental zones were identified using soil and rock types, and so forth) affecting soil Se retention. To a two-step cluster analysis (SI Materials and Methods). Based on the data from each zone, individual parameters were allowed to vary while all other evaluate small-scale soil Se distributions or to test locally relevant variables were held constant at the zonal averages. By using different zones, hypotheses, scale-appropriate models are necessary. we could model the response of soil Se to changes in particular variables Although some effects of climate change on global food security under different environmental conditions. These analyses allowed us to are predictable (e.g., decreased food production resulting from in- identify the most likely mechanism driving soil Se concentrations by com- creased water stress), the predicted widespread reductions in soil Se paring the predictions made by various hypotheses (Table S1) with the caused by climate change were less foreseeable. Changes in other patterns observed in the sensitivity analysis. factors (e.g., specific Se sources, soil properties, soil and rock Each predictive model was also used to predict future (2080–2099) soil Se weathering, and others.) will likely have an additional effect on soil concentrations based on projected climate and SOC changes. Future datasets did not exist for all variables (e.g., clay content); such variables were included Se, but these factors were not analyzed because future projections within the prediction, but their values were identical in the two time points. for soil pH and clay content and spatial information on the con- Although some variables (e.g., sand, silt, and clay content) are not likely to tributions of anthropogenic and natural sources of Se are currently change considerably, changes in other variables, such as soil pH, are likely to unavailable. These variables are likely to have an effect on soil Se result in changes in soil Se concentrations. Therefore, we discuss only po- concentrations. For example, given changes in industrial SO and tential changes in soil Se concentrations resulting from climate change in- NO emissions (55), soil pH will likely increase (56). Increases in pH stead of reporting actual soil Se concentrations. Future predictions were may result in further losses of soil Se concentrations, given that soil retained if the SD of the future prediction was <10% of the mean prediction Se and soil pH are inversely related. Therefore, updated soil Se (i.e., SD < 0.1*mean) or if the three models predicted the same direction predictions are likely to change as additional data become available. (loss or gain) of change (SI Materials and Methods). Only pixels that over- Given the importance of climate–soil interactions on soil Se dis- lapped between the 1980–1999 and 2080–2099 time periods (∼48% of the global land surface) were reported in discussions of future changes. tributions, it is likely that other trace elements with similar retention In addition to predictive analyses, we developed a conceptual model de- mechanisms will experience similar reductions as the result of cli- scribing broad-scale soil Se concentrations based on mechanistic knowledge matic change. Coupled with micronutrient stripping from agricul- gained from the literature and on climate knowledge gained from predictive tural lands (57), predicted losses of total Se in soils indicate that the analyses. This proposed model was evaluated using SEM (i) to test different nutritional quality of food may decrease, thereby increasing the hypotheses proposed to govern soil Se concentrations, (ii)to evaluate simul- worldwide risk of micronutrient deficiency. However, as stated pre- taneously the relative importance of these different hypothesized mecha- viously, total soil Se concentrations are not the only factor de- nisms, and (iii) to evaluate the direct and indirect effects (i.e., mediated termining Se levels in plants. Lower Se levels in soils could effects) of the variables on soil Se concentrations (the direct effects generated potentially compound the problems associated with the decrease in from the SEM analysis are analogous to the univariate sensitivity analysis of the machine-learning models). Although SEM is not predictive, it has advan- the nutritional value of some plants resulting from elevated atmo- tages over the predictive models because it can quantify both the direct and spheric CO concentrations (58). Potential micronutrient losses from indirect effects of all variables more easily, and it was used to help identify agricultural soils could be offset by implementing agricultural prac- important interactions among variables. The SEM was considered a good fit if tices that increase their retention [e.g., organic carbon (OC) ad- the SRMR was ≤0.8 and the CFI was ≥0.95 (43). Only statistically significant (α = justment]; however, such strategies may not increase soil Se in areas 0.05) variables were retained in the SEM analysis. All error intervals presented of increasing aridity, given the importance of AI in governing soil Se represent 95% confidence intervals unless otherwise noted. All statistical concentrations. Where soils cannot be manipulated to increase the procedures were performed using the software packages R (v. 3.3.2, R long-term retention of Se, broad-scale micronutrient fertilization Development Core Team, Vienna), SPSS (v. 22, IBM. Corp., Armonk, NY), and SPSS-Amos (v. 22, IBM Corp., Armonk, NY), and all spatial procedures were may be necessary to maintain an adequate nutrient content in crops. performed using the software packages ArcMap [v. 10.2.2, Environmental Systems Research Institute (ESRI), Redlands, CA] and R. Materials and Methods Total Se concentrations in soils (mg Se/kg soil, reported herein as mg Se/kg; ACKNOWLEDGMENTS. We thank S. Adcock, M. Broadley, E. C. da Silva, Jr., soils were air dried or oven dried) 0–30 cm deep (n = 33,241 samples) were A. Chilimba, K. Dhillon, A. Donald, A. Eqani, L. Guilherme, E. Joy, K. Macey, obtained from Brazil, Canada, China, Europe, Japan, Kenya, Malawi, New A. Meharg, P. Morris, G. Paterson, H. Shen, J. van Ryssen, J. Wilford, and Zealand, South Africa, and the United States (see SI Materials and Methods J. Yanai for providing data; T. Blazina for digitizing the Chinese soil Se data; for dataset details and a discussion about which Se datasets were used, Fig. J. Hernandez for helping us obtain samples from the Rothamsted archive; S8). Samples derived from stream sediments were excluded from this analysis. M. Glendining for providing data for the Broadbalk soil experiment; In addition, we obtained 26 variables describing factors hypothesized to K. Abbaspour, K. Coleman, C. F. Randin, H. F. Satizábal, and R. Siber for input control soil Se concentrations and moderate climate change projections (RCP on methodological approaches; C. Stengel for assistance in analyzing soil samples for Se; R. Jones for comments on an earlier draft of this paper; 6.0 for climate and A1B for SOC data; see Table S1 forvariable descriptionsand U. Beyerle and J. Sedlacek for processing the Climate Model Intercomparison citations). All data within a 1° cell were averaged and represented by a single Project (CMIP) Phase 5 data; the World Climate Research Programme’s Work- value. To minimize the influence of errors and/or outliers within the datasets, ing Group on Coupled Modeling, which is responsible for CMIP; and the pixels containing fewer than five Se data points were removed from the climate modeling groups for producing and making available their model analysis (SI Materials and Methods). The final soil Se dataset consisted of n = output. The US Department of Energy’s Program for Climate Model Diagnosis 1,642 aggregated points. Four techniques for selecting variables [e.g., corre- and Intercomparison provides coordinating support for CMIP and led the de- lations, principal components analysis (PCA), backward elimination modeling, velopment of software infrastructure in partnership with the Global Organi- and RF node purity analyses; see SI Materials and Methods] were used to retain zation for Earth System Science Portals. This work was supported by Swiss National Science Foundation Grants PP00P2_133619 and PP00P2_163747 and the following variables for predictive analysis: AI, clay content, ET, lithology, Eawag, the Swiss Federal Institute of Aquatic Science and Technology. P.S. is pH, precipitation, and SOC at a soil depth of 0–30 cm. Although 16 lithological supported by the Delivering Food Security on Limited Land (DEVIL) project classes were present within the raster dataset, classes that were represented by (UK Natural Environmental Research Council NE/M021327/1) funded by the too few soil Se data points (n < 200) were grouped together instead of being Belmont Forum/Joint Programming Initiative on Agriculture, Food Security deleted (Fig. S4 and see SI Materials and Methods for further discussion). and Climate Change (FACCE-JPI) and by UK Biotechnology and Biological Sci- Predictive modeling was performed using three machine-learning models (one ences Research Council Project BB/L000113/1. Rothamsted Research is grant- RF and two artificial neural network models) (SI Materials and Methods). Each aided by the UK Biotechnology and Biological Sciences Research Council. 2852 | www.pnas.org/cgi/doi/10.1073/pnas.1611576114 Jones et al. 1. Stoltzfus RJ, Mullany L, Black RE (2004) Iron Deficiency Anaemia. Comparative 34. Trabucco A, Zomer R (2010) Global soil water balance geospatial database. CGIAR Quantification of Health Risks: Global and Regional Burden of Disease Attribution Consortium for Spatial Information. Published online, available from the CGIAR-CSI to Selected Major Risk Factors, eds Ezzati M, Lopez AD, Rodgers A, Murray CJL (World GeoPortal at www.cgiar-csi.org. Accessed January 28, 2017. Health Organization, Geneva), Vol 1, pp 163–210. 35. Siebert S, Henrich V, Frenken K, Burke J (2013) Update of the digital global map 2. Adamson P (2004) Vitamin and Mineral Deficiency: A Global Progress Report (Mi- of irrigation areas to version 5 (Rheinische Friedrich-Wilhelms-Universität, Bonn, cronutrient Initiative, UNICEF, Ottawa). Germany and Food and Agriculture Organization of the United Nations, Rome, 3. Fordyce FM (2013) Selenium deficiency and toxicity in the environment. Essentials of Italy). Medical Geology, eds Selinus O, et al. (Springer, London), pp 373–415. 36. Milly PC, Dunne KA, Vecchia AV (2005) Global pattern of trends in streamflow and 4. James L, et al. (1989) Selenium Poisoning in Livestock: A Review and Progress. water availability in a changing climate. Nature 438(7066):347–350. Selenium in Agriculture and the Environment (Soil Science Society of America and 37. Nachtergaele F, et al. (2010) Global land degradation information system (GLADIS), American Society of Agronomy, Madison, WI), pp 123–131. Beta version. (Food and Agriculture Organization of the United Nations, Rome), 5. Chilimba ADC, et al. (2011) Maize grain and soil surveys reveal suboptimal dietary Technical Report 17. selenium intake is widespread in Malawi. Sci Rep 1(72):72. 38. Hartmann J, Moosdorf N (2012) The new global lithological map database GLiM: A 6. Renkema H, et al. (2012) The effect of transpiration on selenium uptake and mobility representation of rock properties at the Earth surface. Geochem Geophys Geosyst in durum wheat and spring canola. Plant Soil 354(1-2):239–250. 13(12):1–37. 7. Winkel LHE, et al. (2015) Selenium cycling across soil-plant-atmosphere interfaces: A 39. Hollmann R, et al. (2013) The ESA climate change initiative: Satellite data records for critical review. Nutrients 7(6):4199–4239. essential climate variables. Bull Am Meteorol Soc 94(10):1541–1552. 8. Sun G-X, et al. (2016) Distribution of soil selenium in China is potentially controlled 40. Broxton PD, Zeng XB, Sulla-Menashe D, Troch PA (2014) A global land cover clima- bydeposition and volatilization? Sci Rep 6:20953. tology using MODIS data. J Appl Meteorol Climatol 53(6):1593–1605. 9. Winkel LHE, et al. (2012) Environmental selenium research: From microscopic pro- 41. Lefsky MA (2010) A global forest canopy height map from the Moderate Resolution cesses to global understanding. Environ Sci Technol 46(2):571–579. Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophys Res 10. Wen H, Carignan J (2007) Reviews on atmospheric selenium: Emissions, speciation and Lett 37(15):L15401. fate. Atmos Environ 41(34):7151–7165. 42. Center for International Earth Science Information Network - CIESIN – Columbia 11. Fernández-Martínez A, Charlet L (2009) Selenium environmental cycling and bio- University (2016) Documentation for the Gridded Population of the World, Version 4 availability: A structural chemist point of view. Rev Environ Sci Biotechnol 8(1):81–110. 12. Neal RH, Sposito G (1991) Selenium mobility in irrigated soil columns as affected by (GPWv4). (NASA Socioeconomic Data and Applications Center, Palisades NY) Avail- organic carbon amendment. J Environ Qual 20(4):808–814. able at sedac.ciesin.columbia.edu/data/collection/gpw-v4/documentation. Accessed 13. Blazina T, et al. (2014) Terrestrial selenium distribution in China is potentially linked January 28, 2017. to monsoonal climate. Nat Commun 5(4717):4717. 43. Hu L, Bentler PM (1999) Cutoff criteria for fit indexes in covariance structure analysis: 14. Lee GK, et al. (2016) The Geochemical Atlas of Alaska, 2016, US Geologic Survey Data Conventional criteria versus new alternatives. Struct Equ Modeling 6(1):1–55. Series 908, p. 25, GIS database available at https://pubs.er.usgs.gov/publication/ds908. 44. Rovira M, et al. (2008) Sorption of selenium(IV) and selenium(VI) onto natural iron Accessed January 28, 2017. oxides: Goethite and hematite. J Hazard Mater 150(2):279–284. 15. Fichtner SS, et al. (2007) Estudo da composição mineral de solos, forragens e tecido 45. Dhillon KS, Dhillon SK (2014) Development and mapping of seleniferous soils in animal de bovinos do município De Rio Verde, Goiás: Cobre, molibdênio e selênio. northwestern India. Chemosphere 99:56–63. Pesqui Agropecu Trop 20(1):1–6. 46. Western AW, Grayson RB, Blöschl G (2002) Scaling of soil moisture: A hydrologic 16. Faria L (2009) Levantamento Sobre Selênio em Solos e Plantas do Brasil e sua Apli- perspective. Annu Rev Earth Planet Sci 30(1):149–180. cação em Plantas Forrageiras. Thesis (Universidade De São Paulo, São Paulo, Brasil). 47. Silver WL, Lugo A, Keller M (1999) Soil oxygen availability and biogeochemistry along 17. Gabos MB, Alleoni LRF, Abreu CA (2014) Background levels of selenium in some se- rainfall and topographic gradients in upland wet tropical forest soils. Biogeochemistry lected Brazilian tropical soils. J Geochem Explor 145:35–39. 44(3):301–328. 18. Carvalho GS (2011) Selênio e Mercúrio em Solos sob Cerrado Nativo. Thesis (Uni- 48. Guo L, Frankenberger WT, Jury WA (1999) Evaluation of simultaneous reduction and versidade Federal de Lavras, Lavras, Brazil). transport of selenium in saturated soil columns. Water Resour Res 35(3):663–669. 19. Smith DB, et al. (2005) Major- and Trace-Element Concentrations in Soils from Two 49. Boulding JR, Ginn JS (2003) Practical Handbook of Soil, Vadose Zone, and Ground- Continental-Scale Transects of the United States and Canada, US Geologic Survey Water Contamination: Assessment, Prevention, and Remediation (CRC, Boca Raton, FL). Open-File Report 2005-1253. Available at https://pubs.usgs.gov/of/2005/1253/pdf/ 50. Fordyce F (2007) Selenium geochemistry and health. Ambio 36(1):94–97. OFR1253.pdf. Accessed January 28, 2017. 51. Ocean Biology Processing Group (2003) MODIS Aqua Level 3 Global Monthly Mapped 20. Friske P, et al. (2014) Soil Geochemical, Mineralogical, Radon and Gamma Ray Spec- 4 km Chlorophyll a. Ver. 6. (Physical Oceanography Distributed Active Archive Center, trometric Data from the 2007 North American Soil Geochemical Landscapes Project in Pasadena, CA). New Brunswick, Nova Scotia, and Prince Edward Island; Geological Survey of 446 52. Taylor KE, Stouffer RJ, Meehl GA (2012) An overview of Cmip5 and the experiment Canada, Open File 6433 (revised). Available at http://geogratis.gc.ca/api/en/nrcan- design. Bull Am Meteorol Soc 93(4):485–498. rncan/ess-sst/fa531017-b8bf-50c0-b7e4-a7764361415b.html. Accessed January 28, 2017. 53. Long JA, et al. (2015) Severe selenium depletion in the Phanerozoic oceans as a factor 21. Friske P, et al. (2013) Soil Geochemical, Radon and Gamma Ray Spectrometric Data in three global mass extinction events. Gondwana Res 36:209–218. from the 2008 and 2009 North American Soil Geochemical Landscapes Project Field 54. Yu T, et al. (2014) The origin and geochemical cycle of soil selenium in a Se-rich area Surveys, Open File 7334. Available at http://geoscan.nrcan.gc.ca/starweb/geoscan/ of China. J Geochem Explor 139:97–108. servlet.starweb?path=geoscan/fullf.web&search1=R=292514. Accessed January 28, 2017. 55. Intergovernmental Panel on Climate Change (2014) Climate Change 2014: Synthesis 22. Rawlins B, et al. (2012) The Advanced Soil Geochemical Atlas of England and Wales. (British Geological Survey, Keyworth, UK). Report, eds IPCC Core Writing Team, Pachauri RK, Meyer LA (IPCC, Geneva). 23. Reimann C, et al. (2014) Distribution of elements/parameters in agricultural and 56. Kirk GJD, Bellamy PH, Lark RM (2010) Changes in soil pH across England and Wales in grazing land soil in Europe. Chemistry of Europe’s Agricultural Soils. Part A: response to decreased acid deposition. Glob Change Biol 16(11):3111–3119. Methodology and Interpretation of the GEMAS Data Set, eds Reimann C, Birke M, 57. Jones DL, et al. (2013) Review: Nutrient stripping: The global disparity between food Demetriades A, Filzmoser P, O’Connor P (Bundesanstalt für Geowissenschaften und security and soil nutrient stocks. J Appl Ecol 50(4):851–862. Rohstoffe, Hannover, Germany), pp 103–474. 58. Myers SS, et al. (2014) Increasing CO threatens human nutrition. Nature 510(7503): 24. Yamada H, Kamada A, Usuki M, Yanai J (2009) Total selenium content of agricultural 139–142. soils in Japan. Soil Sci Plant Nutr 55(5):616–622. 59. Oldfield J (2002) Selenium World Atlas. (Selenium-Tellurium Development Associa- 25. Maskall JE, Thornton I (1989) The mineral status of Lake Nakuru National Park, Kenya - a tion, Grimbergen, Belgium). Updated edition. reconnaissance survey. Afr J Ecol 27(3):191–200. 60. Satizábal MH, Pérez-Uribe A (2007) Relevance metrics to reduce input dimensions in 26. Joy EJM, et al. (2015) Soil type influences crop mineral composition in Malawi. Sci artificial neural networks. Proceedings of the 17th International Conference on Ar- Total Environ 505:587–595. tificial Neural Networks, September 9–13, eds Sá JMd, Alexandre L, Duch W, Mandic D 27. Martin A, et al. (2016) The regional geochemical baseline soil survey of southern New (Springer, Berlin), pp 39-48. Zealand: Design and initial interpretation. J Geochem Explor 167:70–82. 61. Carpita M, Sandri M, Simonetto A, Zuccolotto P (2013) Football mining with R. Data 28. Zheng C (1994) Atlas of Soil Environmental Background Value in the People’s Mining Applications with R, eds Zhao Y, Cen Y (Academic, Waltham, MA), pp Republic of China (China Environmental Science Press, Beijing), p 195. 397–433. 29. ARC-ISCW (2005) Overview of the Agricultural Natural Resources of South Africa: Report 62. Nocedal J, Wright S (2006) Numerical Optimization (Springer Science & Business No. GW/A/2004/13. (ARC-Institute for Soil, Climate and Water, Pretoria, South Africa). Media, New York), 2nd Ed. 30. Greve P, Seneviratne SI (2015) Assessment of future changes in water availability and 63. Knights JS, Zhao FJ, Spiro B, McGrath SP (2000) Long-term effects of land use and aridity. Geophys Res Lett 42(13):5493–5499. fertilizer treatments on sulfur cycling. J Environ Qual 29(6):1867–1874. 31. Stackhouse PW, Jr, et al. (2006) Fast longwave and shortwave radiative flux 64. Hewitt A, Reynolds C (1990) Dissolution of metals from soils and sediments with a (FLASHFlux) products from CERES and MODIS measurements. Proceedings of the 12th microwave-nitric acid digestion technique. Spectrosc. 11(5):187–192. Conference on Atmospheric Radiation,July10–14, Madison, WI, (American Meteo- 65. Binstock DA, et al. (1990) Validation of Methods for Determining Elements in Solid rological Society, Boston), pp 1–6. Waste by Microwave Digestion. Waste Testing and Quality Assurance, ed Friedman D 32. Hengl T, et al. (2014) SoilGrids1km–global soil information based on automated (ASTM International, Philadelphia), Vol 2, pp 259–270. mapping. PLoS One 9(8):e105992. 66. Ramankutty N, Evan AT, Monfreda C, Foley JA (2008) Farming the planet: 1. Geo- 33. Gottschalk P, et al. (2012) How will organic carbon stocks in mineral soils evolve under graphic distribution of global agricultural lands in the year 2000. Global future climate? Global projections using RothC for a range of climate change sce- narios. Biogeosciences 9(8):3151–3171. Biogeochem Cycles 22(1):GB1003. Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2853 ENVIRONMENTAL SCIENCES http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Proceedings of the National Academy of Sciences of the United States of America Pubmed Central

Selenium deficiency risk predicted to increase under future climate change

Proceedings of the National Academy of Sciences of the United States of America , Volume 114 (11) – Feb 21, 2017

Loading next page...
 
/lp/pubmed-central/selenium-deficiency-risk-predicted-to-increase-under-future-climate-84D07UsXCg

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
Pubmed Central
ISSN
0027-8424
eISSN
1091-6490
DOI
10.1073/pnas.1611576114
Publisher site
See Article on Publisher Site

Abstract

Selenium deficiency risk predicted to increase under future climate change a a b c a,d e b Gerrad D. Jones , Boris Droz , Peter Greve , Pia Gottschalk , Deyan Poffet , Steve P. McGrath , Sonia I. Seneviratne , f a,d,1 Pete Smith , and Lenny H. E. Winkel a b Eawag, Swiss Federal Institute of Aquatic Science and Technology, 8600 Duebendorf, Switzerland; Institute for Atmospheric and Climate Science, ETH c d Zurich, 8092 Zurich, Switzerland; Potsdam Institute for Climate Impact Research, 14473 Potsdam, Germany; Institute of Biogeochemistry and Pollutant Dynamics, ETH Zurich, 8092 Zurich, Switzerland; Department of Sustainable Soils and Grassland Systems, Rothamsted Research, Harpenden AL5 2JQ, United Kingdom; and Institute of Biological and Environmental Sciences, School of Biological Sciences, University of Aberdeen, Aberdeen AB24 3UU, United Kingdom Edited by Jerome Nriagu, University of Michigan, Ann Arbor, MI, and accepted by Editorial Board Member David W. Schindler January 6, 2017 (received for review July 15, 2016) Deficiencies of micronutrients, including essential trace elements, were (i) to test hypothesized drivers of soil Se concentrations, (ii)to affect up to 3 billion people worldwide. The dietary availability of predict global soil Se concentrations quantitatively, and (iii)to trace elements is determined largely by their soil concentrations. quantify potential changes in soil Se concentrations resulting from Until now, the mechanisms governing soil concentrations have been climate change. To achieve these objectives, several regional- to evaluated in small-scale studies, which identify soil physicochemical continental-scale datasets reporting total soil Se concentrations [n = properties as governing variables. However, global concentrations of 33,241 data points (5, 14–29); see SI Materials and Methods for trace elements and the factors controlling their distributions are dataset details] and 26 environmental variables describing climate, virtually unknown. We used 33,241 soil data points to model recent soil physicochemical properties, irrigation, water stress, erosion, (1980–1999) global distributions of Selenium (Se), an essential trace runoff, land use, soil type, lithology, bedrock depth, vegetation/ element that is required for humans. Worldwide, up to one in seven canopy characteristics, and population density (30–42) (see Table people have been estimated to have low dietary Se intake. Contrary S1 for details of variables) were assessed to make global predictions to small-scale studies, soil Se concentrations were dominated by of soil Se concentrations for recent (1980–1999) and future (2080– climate–soil interactions. Using moderate climate-change scenarios 2099) periods. Predictions were made using three machine-learning for 2080–2099, we predicted that changes in climate and soil organic tools: one randomForest (RF) model and two artificial neural carbon content will lead to overall decreased soil Se concentrations, network models, herein referred to as “predictive models.” Addi- particularly in agricultural areas; these decreases could increase the tionally, structural equation modeling (SEM) was used to evaluate prevalence of Se deficiency. The importance of climate–soil interac- potential mechanisms independently and to quantify complex in- tions to Se distributions suggests that other trace elements with sim- teractions between Se and the relevant predictor variables. ilar retention mechanisms will be similarly affected by climate change. Results and Discussion selenium soils global distribution prediction climate change | | | | After variable selection, seven variables were retained and were considered the most important factors controlling soil Se con- icronutrients are essential for maintaining human health, centrations: the aridity index (AI, unitless) [i.e., the ratio of po- Mand although they are needed in only trace amounts, de- tential evapotranspiration (PET, mm/d) to precipitation (mm/d)]; ficiencies reportedly affect 3 billion people worldwide (1, 2). One such micronutrient is selenium. Inadequate dietary Se intake af- Significance fects up to 1 in 7 people and is also known to affect livestock health adversely (3, 4). Because dietary Se intake depends largely on Se The trace element selenium is essential for human health and is content in soil and bioavailability to crops (5–7), understanding the required in a narrow dietary concentration range. Insufficient mechanisms driving soil concentrations and predicting global dis- selenium intake has been estimated to affect up to 1 billion tributions could help prevent Se deficiency (8). However, global people worldwide. Dietary selenium availability is controlled by soil Se concentrations and the factors affecting Se distributions are soil–plant interactions, but the mechanisms governing its broad- largely unknown (9). Apart from soils, Se is present in all other scale soil distributions are largely unknown. Using data-mining environmental compartments [i.e., the lithosphere, hydrosphere, techniques, we modeled recent (1980–1999) distributions and biosphere, and atmosphere (9)], which all play a role in global Se identified climate–soil interactions as main controlling factors. biogeochemical cycling and distribution (7, 10). Furthermore, using moderate climate change projections, we The factors driving soil Se concentrations [e.g., increased sorption predicted future (2080–2099) soil selenium losses from 58% of with decreased pH and soil reduction potential (Eh) and increased modeled areas (mean loss = 8.4%). Predicted losses from crop- clay and soil organic carbon (SOC) content; see Table S1 and refs. 7 lands were even higher, with 66% of croplands predicted to lose and 11 for a review of the previously reported drivers of soil Se] 8.7% selenium. These losses could increase the worldwide have been evaluated primarily through small-scale experimentation prevalence of selenium deficiency. (e.g., soil columns; see ref. 12); however, broad-scale distributions cannot be inferred from such studies. For example, soils in small- Author contributions: G.D.J., B.D., and L.H.E.W. designed research; G.D.J., B.D., D.P., and L.H.E.W. scale experiments are often manipulated [e.g., by carbon amend- performed research; G.D.J., B.D., P. Greve, P. Gottschalk, S.P.M., S.I.S., P.S., and L.H.E.W. ments (12)] to achieve desired conditions, obscuring the natural contributed data/soil samples; G.D.J., B.D., D.P., and L.H.E.W. analyzed data; and G.D.J., B.D., D.P., and L.H.E.W. wrote the paper with technical input from all authors. processes that may influence Se retention capacity. Additionally, climate variables, which likely affect soil Se concentrations directly The authors declare no conflict of interest. as a source (e.g., deposition; see refs. 8 and 13) or indirectly by This article is a PNAS Direct Submission. J.N. is a Guest Editor invited by the Editorial Board. affecting soil retention of Se (e.g., sorption), are ignored in small- scale experiments. Therefore, to predict the global distributions, Freely available online through the PNAS open access option. broad-scale analyses of soil Se drivers are essential. To whom correspondence should be addressed. Email: lwinkel@ethz.ch. Here we report on the influence of soil and climate variables on This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. worldwide Se distributions in soils 0–30 cm deep. Our objectives 1073/pnas.1611576114/-/DCSupplemental. 2848–2853 | PNAS | March 14, 2017 | vol. 114 | no. 11 www.pnas.org/cgi/doi/10.1073/pnas.1611576114 world. In sensitivity analyses, soil Se increased with increases in clay content and with decreases in soil pH (Fig. 2 and Figs. S3 and S4), both of which are known to increase soil Se sorption (7, 44). Al- though soil Se is known to partition/complex with organic matter (7), soil Se was affected only weakly by changes in SOC (Figs. S1, S3,and S4 and Table S2). Furthermore, changes in lithological classes resulted in negligible changes in soil Se when other variables were held constant (Fig. S4). Although lithology was of minor im- portance in this study, we recognize that it can influence soil Se concentrations at local scales (45). Climate Effects on Soil Se. Climate variables (i.e., AI, precipitation, and ET) were dominant factors driving soil Se concentrations, likely because they control leaching from soils, and observed pat- terns within the sensitivity and SEM analyses for all climate vari- ables are consistent with this hypothesis. High precipitation and AI negatively affected soil Se concentrations, whereas ET positively affected soil Se concentrations (Fig. 2, Figs. S3 and S4,and Table S2). Although AI (i.e., the ratio of PET to precipitation) and precipitation are inversely related, both variables exerted negative effects on soil Se, suggesting that separate mechanisms drive these patterns. Although precipitation increases the transport of dis- solved Se species in soil solution by increasing vadose zone flow, AI likely affects leaching by controlling soil redox conditions and thus Se speciation, sorption, and mobility. It has been reported that as AI increases (i.e., PET increases relative to precipitation), soils become drier (46), resulting in more oxidizing soil conditions (47). Oxidized Fig. 1. Summary of the processes governing soil Se concentrations. Domi- Se species (e.g., oxyanions) are more soluble and mobile than re- nant processes (and bulleted examples) governing soil Se concentrations are duced species (e.g., selenides) (7, 11, 12, 48). Therefore, soil drying indicated. Text colored in red, green, and blue indicates processes affecting soil Se losses, retention, and sources/supplies, respectively. Factors re- likely increases the presence of oxidized/mobile soil Se species, sponsible for increases (+) and/or decreases (−) in soil Se as well as processes which can be leached during subsequent rain events. not explicitly examined in our analysis (*) are indicated. Soil drying likely increases Se mobility but also can reduce soil Se transport (although these processes likely occur at different time scales). Leaching is driven by the ratio of ET to precipitation clay content (%); evapotranspiration (ET, mm/d); lithology; pH; (also known as the “evaporative index,” EI) (49). As previously precipitation (mm/d); and SOC (0–30 cm depth, tons of C/ha). mentioned, precipitation increases the transport of Se through the With these variables, the accuracy of the predictive model was vadose zone, but as ET increases relative to precipitation, more 2 2 high (average R = 0.67, average cross-validation R = 0.49, n = moisture is removed from the soil column. This removal of moisture 1,000 iterations for each model) (Fig. S1), and the precision was high [based on a low SD of the modeled prediction (0.032 mg Se/kg) relative to the mean (0.35 mg Se/kg)] (Fig. S2). For the SEM 1.2 1.2 analysis, both the standardized root mean squared residual A B 1.0 (SRMR) (0.043) and the comparative fit index (CFI) (0.962) in- Zone 1 1.0 Zone 2 dicated a good fit between the observed and modeled data (43). 0.8 0.8 Zone 3 All variables retained within the SEM analysis were statistically 0.6 0.6 significant (i.e., P < 0.05) (Fig. S3), and, based on the predictive 0.4 0.4 model sensitivity analyses and the SEM regression weights, the 0.2 0.2 modeling results corroborated each other strongly (see following 0.0 0.0 section), suggesting that processes driving soil Se concentrations 048 12 468 10 Aridity Index (AI) [unitless] pH were described accurately. Because the data were averaged on a 1° 1.2 1.2 scale, this modeling approach likely captures the broad-scale C 4.2 D 0.5 pH Evapotranspiraon 4.8 1.0 1.0 1.0 (ET) [mm/d] mechanisms but potentially misses important local-scale factors. 5.3 1.5 5.9 2.0 0.8 0.8 Soil Se concentrations were determined largely by interactions 6.4 2.5 ET = precipitaon between climate and soil variables (Figs. 1 and 2, Figs. S1, S3,and 7.0 3.0 0.6 0.6 7.5 3.5 S4,and Table S2). In the SEM, AI and precipitation had the 4.0 8.1 0.4 0.4 4.5 8.6 greatest direct and indirect effect, respectively, on soil Se concen- 9.2 5.0 0.2 0.2 trations (Table S2). Basedonaveragedrelative importancefrom 0.0 0.0 the predictive models (Fig. S1), AI was the most important pre- 0.1 1 10 0123456 dictor (100 ± 0.3%) followed by pH (60 ± 0.7%), precipitation (58 ± Aridity Index (AI) [unitless] Precipitaon [mm/d] 1%), ET (50 ± 0.8%), clay content (45 ± 0.4%), lithology (33 ± Fig. 2. Univariate and bivariate sensitivity analyses of the predictive models. 0.5%), and SOC (29 ± 0.5%). Sensitivity analyses were performed (A and B) The independent effects of AI (A)and precipitation (B)weremod- on these variables to determine if the mechanisms driving soil Se eled by holding all other variables constant at the zonal averages as defined by concentrations changed in different zones represented by different thetwo-stepclustering. (C and D) Similarly, bivariate interactions between AI environmental conditions (see SI Materials and Methods for a de- and clay (C) and between precipitation and ET (D) are illustrated. These pa- scription). The soil Se patterns were similar between different zones, rameters were allowed to vary between the minimum and maximum observed suggesting that Se drivers were consistent regardless of the envi- value while all other variables were held constant at the mean value of the ronment (Fig. 2 and Fig. S4). This result suggests that the models entire dataset (n = 1,642). The dotted line in D indicates the conditions in can be used to predict soil Se concentrations in other regions of the which ET = precipitation. Other bivariate interactions are presented in Fig. S3. Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2849 Modeled soil selenium [mg/kg] ENVIRONMENTAL SCIENCES reduces the vadose zone flow, which in turn reduces Se mass sensitivity and SEM analyses suggest that the direct (i.e., non- transport. In theory, when ET and precipitation are equal, leaching mediated) effect of precipitation drives Se leaching from soils (Fig. 2 should be negligible. Although ET clearly dampens the negative and Figs. S3 and S4), thus explaining why low values resulted in high effects of precipitation in bivariate sensitivity analyses, a negative soil Se, even though the net effect is positive (Table S2). In bivariate relationship existed between precipitation and modeled Se despite sensitivity analyses, soil Se concentrations exceeded these values an ET:precipitation ratio of 1 (Fig. 2). This trend potentially could when low AI was modeled with low pH (1.12 mg Se/kg), when low be explained by plant Se uptake, which has been reported to precipitation was modeled with high clay content (0.86 mg Se/kg), increase with ET (6). Therefore, in addition to its positive indi- and when high clay content was modeled with low pH (0.56 mg Se/kg) rect effect by reducing leaching, ET also may have a direct negative (Fig. 2 and Fig. S3). Although Se concentrations were typically en- effect on soil Se by increasing plant uptake. This direct negative hanced in low-AI or low-precipitation environments, both variables effect was observed in the SEM analysis, but the relationship was could suppress the effects of other variables in high-AI or high- not statistically significant and therefore was removed. Although precipitation environments (Fig. 2 and Fig. S3). These results this negative effect may exist, it appears to be less important than demonstrate the dependence of soil Se concentrations on soil– the role ET plays in reducing Se leaching. Given the importance of climate interactions. Based on these analyses, low-Se soils are climate variables in governing soil Se concentrations, the observed most likely to occur in arid environments and in areas with high patterns between AI, precipitation, ET, and modeled soil Se con- pH and low clay content. Conversely, areas with low to moderate centrations strongly suggest that changes in climate will result in precipitation but relatively low aridity (e.g., cool and moist climates) changes in soil Se concentrations in time and space. and high clay content are likely to have higher soil Se concentrations. Climate–Soil Interactions and Soil Se. Although the direct effects of Predicted Global Soil Se Distributions. Global predictions were precipitation (i.e., leaching) were moderate, its indirect effects (i.e., made using models trained largely using temperate/midlatitude those mediated through other variables) were approximately three- datasets (Fig. S2). Although the available data adequately described fold larger (Table S2). Precipitation is known to affect soil forma- similar regions, data from tropical, extremely arid, and polar regions tion (i.e., pedogenesis), and in our analyses it strongly affected AI, were almost entirely absent (to the best of our knowledge, no broad- pH, ET, and clay content, which subsequently affect soil Se re- scale soil geochemical surveys are available from these regions). As tention (Fig. S3 and Table S2). Although there was a negative direct a result, predictions that were made for environments that fell effect between precipitation and soil Se, the sum of direct and in- outside our dataset’s domain were excluded (Fig. S5). Therefore, Se direct effects resulted in precipitation having a net positive effect predictions for 1980–1999 were retained for only 70% of land sur- (Table S2). Thus it is important to examine both direct and indirect faces. The majority of croplands and rangelands, which are areas of effects, because interpreting only total effects can lead to erroneous primary interest, given that soil Se concentrations and bioavailability conclusions about the mechanisms driving spatial patterns. in these regions largely drive the Se status in humans and livestock, In bivariate sensitivity analyses, both synergistic and antagonistic fall largely within the retained areas. interactions were observed and were strongest between aridity, Based on predictive models, the global mean soil Se concen- precipitation, clay content, and pH. In univariate sensitivity anal- tration for 1980–1999 was 0.322 ± 0.002 mg Se/kg (Fig. 3), similar yses, when all other variables were held constant, modeled soil Se to reported values (mean = 0.4 mg Se/kg; typical range 0.01–2mg concentrations were highest under low AI (0.83 mg Se/kg), low Se/kg) (50). Using this estimate, ∼13.1 million metric tons of Se precipitation (0.65 mg Se/kg), low pH (0.51 mg Se/kg), and relatively are stored in the top 30 cm of soil within the predicted area [i.e., 7 2 high clay content (0.47 mg Se/kg) (Fig. 2 and Fig. S4). It is important ∼70% of world’sland surface (1.04 × 10 km ); see SI Materials to note that, as long as PET is sufficiently low, environments with and Methods for the calculation]. Compared with other regions, low precipitation can have low AI values also. Furthermore, predicted soil Se concentrations were generally higher (typically Modeled Soil Se 1980-1999 <0.1 mg/kg 0.1-0.2 mg/kg 0.2-0.3 mg/kg 0.3-0.4 mg/kg 0.4-0.5 mg/kg >0.5 mg/kg Avg. = 0.32 mg/kg % change in Fig. 3. Geographical representation of the pre- soil Se 1980-1999 dictive modeling on a 1° scale. Maps illustrate the to 2080-2099 modeled soil Se concentrations (1980–1999) (A)and <-10% percentage change in soil Se concentrations be- -10--2.5% tween recent and future (2080–2099) conditions (B) -2.5-2.5% as a function of projected changes in climate (RCP6.0 2.5-10% scenario) and SOC content (ECHAM5-A1B scenario). >10% Predictions represent the average of the predictive Changes due to projected changes Avg. = -4.3% models and were based on the AI, soil clay content, in climate variables and SOC ET, lithology, pH, precipitation, and SOC. 2850 | www.pnas.org/cgi/doi/10.1073/pnas.1611576114 Jones et al. >0.2 mg Se/kg) (Fig. 3) in temperate and northern latitudes. In amount that is ∼20–30% of the total estimated Se mass that is cy- wet equatorial regions, concentrations were typically 0.3–0.5 mg Se/kg. cled yearly through the troposphere [i.e., 13,000–19,000 tons (as- Relatively low-Se soils (<0.2 mg Se/kg) were predicted for 15% of sumed to be metric tons)/y (10)], although our estimate is only for modeled areas and were restricted primarily to arid and semiarid 48% of the land surface. Our modeling approach is not a mass regions in Argentina, Australia, Chile, China, southern Africa, and the balance model, so Se fate could not be investigated. Nonetheless, southwestern United States. In some of these countries, low Se changes in Se concentrations in other environmental compartments content in crops and livestock has been reported (3), but it is im- are known from the past. For example, marine Se concentrations portant to note that many factors contribute to low Se content in throughout various periods of the Phanerozoic eon have been 1.5–2 plants (e.g., plant uptake pathways, soil Se speciation, and the orders of magnitude lower than current oceanic concentrations (53). 7 2 abundance of competing ions such as sulfate) (7). Based on areas with future predictions (7.19 × 10 km ), 58% of lands were predicted to lose soil Se (i.e., ΔSe less than −2.5%; Over- and Underpredictions of the Model. In an attempt to identify mean change = −8.4%); 20% were predicted to undergo minor potential missing variables, we examined the residuals of the changes (i.e., −2.5% < ΔSe < 2.5%; mean change = −0.3%); and predictive models. Spatial patterns of any missing variable should 22% were predicted to gain soil Se (i.e., ΔSe > 2.5%; mean change match those of the residuals (Fig. S2). Overall, the models = 5.7%)asa result of changesin climate and SOC(Fig. 3).Pre- underpredicted soil Se concentrations (average residual = −0.036 ± dicted soil Se losses were driven largely by changes in AI, whereas 0.009 mg Se/kg), suggesting that Se sources may be missing from soil Se gains were driven largely by changes in precipitation and the model. On a localized level, soil Se concentrations appear to SOC (Fig. S1). Compared with the total land surface, croplands be underpredicted in regions adjacent to regions of high marine were expected to lose more soil Se. Based on future predictions for productivity (e.g., western Alaska, western Ireland, western Norway, 6 2 croplands (7.55 × 10 km ), 66% of lands were predicted to lose western England, and Wales) (Fig. S2) (51). Marine environments soil Se (ΔSe less than −2.5%; mean change = −8.7%); 15% were are thought to increase soil Se concentrations via wet deposition predicted to undergo minor changes (–2.5% < ΔSe < 2.5%; mean (10, 13), and atmospheric deposition of Se thus could explain some change = −0.4%); and 19% were predicted to gain soil Se (ΔSe > of the model’s underprediction. However, global spatial data do 2.5%; mean change = 7.3%) (Fig. 3 and Fig. S6). Global pasture not exist for Se deposition and thus could not be analyzed. We lands also were predicted to lose soil Se, but to a lesser extent than included population density as a potential proxy for anthropogenic croplands, suggesting that Se deficiency in livestock could increase. emissions, but this was one of the least important variables in 7 2 Based on future predictions for pasture lands (2.55 × 10 km ), the variable selection procedure. We evaluated a wide variety of 61% of lands were predicted to lose soil Se (ΔSe less than −2.5%; qualitative factors [e.g., specific agricultural soil types (e.g., paddy mean change = −8.0%); 19% were predicted to undergo minor soils), specific sedimentary depositional environments (e.g., glacial changes (−2.5% < ΔSe < 2.5%; mean change = −0.4%); and 21% deposits), coal power plants, carbonaceous shale deposits, and were predictedto gainsoilSe(ΔSe > 2.5% mean change = 8.2%) others] that may affect soil Se distributions; however, we found no (Fig. 3 and Fig. S6). Areas with notable losses (i.e., ΔSe less than consistent discernable link between these variables and the broad- −10%) include croplands of Europe and India, pastures of scale distribution of the model residuals. China, Southern Africa, and southern South America, and the Despite the underprediction, overall patterns of modeled soil southwestern United States (Fig. 3 and Fig. S6). Areas of notable Se distribution match the actual distribution quite closely (Fig. S2), gain (ΔSe > 10%) are scattered across Australia, China, India, and 71% of predicted values were within ±0.05 mg Se/kg of the and Africa (Fig. 3 and Fig. S6). observed value, indicating that a majority of the predicted data were relatively accurate. Furthermore, the sensitivity analysis and Temporal Changes in Soil Se. Although our analysis indicates that SEM trends closely match hypothesized mechanisms governing soil future climate change will likely result in widespread changes in Se concentrations reported in the literature (Fig. 2 and Figs. S3 and soil Se, it does not indicate rates of change. To understand the S4). This finding suggests that the models are largely accurate and temporal changes in soil Se concentrations better, we analyzed for capture the dominant processes controlling broad-scale Se distri- soil Se and SOC in a subset of agricultural samples collected from butions. Nevertheless, future studies could include additional pre- the Broadbalk Experiment (Rothamsted, United Kingdom) be- dictor variables, especially those that are currently unavailable, to tween 1865 and 2010. Soil samples were taken from a control plot provide better estimates of broad-scale soil Se. Furthermore, to (unfertilized since 1843) and two “wilderness” plots (a maintained overcome some of this study’s limitations, predictions could be grassland and woodland), which were converted from the control made on more local/regional scales using higher resolution data. plot in 1882 (SI Materials and Methods,and Table S3). The ac- cumulation of Se in soil through time was statistically greater in Modeled Losses of Future Soil Se. The interactions between pre- the wilderness plots than in the control plot [one-way analysis of cipitation and other soil/climate variables strongly suggest that covariance (ANCOVA); year: F(1, 31) = 20.7, P < 0.01; plot F(2, climate changes could drive changes in soil Se concentrations. To 31) = 17.3, P < 0.01]. When controlling for SOC, however, there assess the influence of changes in climate and SOC, soil Se was were no statistical differences between the plots (ANCOVA; modeled for 2080–2099 using moderate climate change scenarios SOC: F(1, 30) = 10.7, P < 0.01; year: F(1, 30) = 6.0, P < 0.05; plot: [Representative Concentration Pathways (RCP) 6.0 for pre- F(2, 30) = 3.2, P > 0.05), indicating that increases in SOC were cipitation, AI, and ET (52) and European Centre/Hamburg driving soil Se accumulation in the model (Fig. S7), as is consistent Model (ECHAM) 5-A1B for SOC (33)]. Other climate scenarios with the results of the future modeling (Fig. S1). Natural changes in (e.g., RCP 8.5) were not used because SOC data were available soil Se concentrations previously have been hypothesized to occur only for A1B scenarios, which are most similar to RCP 6.0. over longer time scales (e.g., hundreds to thousands of years) (54); Future predictions were made for the entire globe, but, based on however, given that SOC and Se began to accumulate on these plots the filtering criteria used (SI Materials and Methods), predictions immediately after conversion, these results suggest that changes in were retained for ∼48% of theglobal land area. Basedonthese soil Se will follow environmental changes rapidly, perhaps on an pixels alone, soil Se concentrations were predicted to drop by 4.3% annual to decadal time scale. Between ∼1880 and 1980, soil Se on average, from 0.331 ± 0.003 mg Se/kg in 1980–1999 to 0.316 ± 0.002 mg Se/kg in 2080–2099, as a result of changes in climate and concentrations increased by ∼15, 35, and 60% on the control, SOC concentrations (Fig. 3). For soil at a depth of 0–30 cm, this loss grassland, and woodland plots, respectively (Fig. S7), indicating that corresponds to ∼403,763 tons of Se over 100 y, or 4,037.6 tons of Se the magnitude of changes predicted to occur by the end of the 21st lost peryear(see SI Materials and Methods for the calculation), an century is plausible. The rates of change in soil Se concentrations Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2851 ENVIRONMENTAL SCIENCES model was iterated 1,000 times using 90% of the data for model training and following environmental perturbations is largely unstudied and 10% of the data for cross-validation for each iteration. The training and cross- should be evaluated further to understand soil Se dynamics better. validation data were chosen at random for each iteration. The model predictions were averaged to estimate recent (1980–1999) global soil Se concentrations; Outlook however, predictions were considered valid only if the environmental parame- One of our aims was to identify the broad-scale mechanisms ters for each pixel fit within the domain of the observed data (Fig. S5). governing soil Se retention. Therefore, at a 1° resolution, the data Sensitivity analyses were performed during each iteration to investigate used are likely too coarse to evaluate or identify the influence of the independent effect of each variable on modeled soil Se concentrations. many small- to regional-scale factors (e.g., local sources, specific Based on all input variables, three environmental zones were identified using soil and rock types, and so forth) affecting soil Se retention. To a two-step cluster analysis (SI Materials and Methods). Based on the data from each zone, individual parameters were allowed to vary while all other evaluate small-scale soil Se distributions or to test locally relevant variables were held constant at the zonal averages. By using different zones, hypotheses, scale-appropriate models are necessary. we could model the response of soil Se to changes in particular variables Although some effects of climate change on global food security under different environmental conditions. These analyses allowed us to are predictable (e.g., decreased food production resulting from in- identify the most likely mechanism driving soil Se concentrations by com- creased water stress), the predicted widespread reductions in soil Se paring the predictions made by various hypotheses (Table S1) with the caused by climate change were less foreseeable. Changes in other patterns observed in the sensitivity analysis. factors (e.g., specific Se sources, soil properties, soil and rock Each predictive model was also used to predict future (2080–2099) soil Se weathering, and others.) will likely have an additional effect on soil concentrations based on projected climate and SOC changes. Future datasets did not exist for all variables (e.g., clay content); such variables were included Se, but these factors were not analyzed because future projections within the prediction, but their values were identical in the two time points. for soil pH and clay content and spatial information on the con- Although some variables (e.g., sand, silt, and clay content) are not likely to tributions of anthropogenic and natural sources of Se are currently change considerably, changes in other variables, such as soil pH, are likely to unavailable. These variables are likely to have an effect on soil Se result in changes in soil Se concentrations. Therefore, we discuss only po- concentrations. For example, given changes in industrial SO and tential changes in soil Se concentrations resulting from climate change in- NO emissions (55), soil pH will likely increase (56). Increases in pH stead of reporting actual soil Se concentrations. Future predictions were may result in further losses of soil Se concentrations, given that soil retained if the SD of the future prediction was <10% of the mean prediction Se and soil pH are inversely related. Therefore, updated soil Se (i.e., SD < 0.1*mean) or if the three models predicted the same direction predictions are likely to change as additional data become available. (loss or gain) of change (SI Materials and Methods). Only pixels that over- Given the importance of climate–soil interactions on soil Se dis- lapped between the 1980–1999 and 2080–2099 time periods (∼48% of the global land surface) were reported in discussions of future changes. tributions, it is likely that other trace elements with similar retention In addition to predictive analyses, we developed a conceptual model de- mechanisms will experience similar reductions as the result of cli- scribing broad-scale soil Se concentrations based on mechanistic knowledge matic change. Coupled with micronutrient stripping from agricul- gained from the literature and on climate knowledge gained from predictive tural lands (57), predicted losses of total Se in soils indicate that the analyses. This proposed model was evaluated using SEM (i) to test different nutritional quality of food may decrease, thereby increasing the hypotheses proposed to govern soil Se concentrations, (ii)to evaluate simul- worldwide risk of micronutrient deficiency. However, as stated pre- taneously the relative importance of these different hypothesized mecha- viously, total soil Se concentrations are not the only factor de- nisms, and (iii) to evaluate the direct and indirect effects (i.e., mediated termining Se levels in plants. Lower Se levels in soils could effects) of the variables on soil Se concentrations (the direct effects generated potentially compound the problems associated with the decrease in from the SEM analysis are analogous to the univariate sensitivity analysis of the machine-learning models). Although SEM is not predictive, it has advan- the nutritional value of some plants resulting from elevated atmo- tages over the predictive models because it can quantify both the direct and spheric CO concentrations (58). Potential micronutrient losses from indirect effects of all variables more easily, and it was used to help identify agricultural soils could be offset by implementing agricultural prac- important interactions among variables. The SEM was considered a good fit if tices that increase their retention [e.g., organic carbon (OC) ad- the SRMR was ≤0.8 and the CFI was ≥0.95 (43). Only statistically significant (α = justment]; however, such strategies may not increase soil Se in areas 0.05) variables were retained in the SEM analysis. All error intervals presented of increasing aridity, given the importance of AI in governing soil Se represent 95% confidence intervals unless otherwise noted. All statistical concentrations. Where soils cannot be manipulated to increase the procedures were performed using the software packages R (v. 3.3.2, R long-term retention of Se, broad-scale micronutrient fertilization Development Core Team, Vienna), SPSS (v. 22, IBM. Corp., Armonk, NY), and SPSS-Amos (v. 22, IBM Corp., Armonk, NY), and all spatial procedures were may be necessary to maintain an adequate nutrient content in crops. performed using the software packages ArcMap [v. 10.2.2, Environmental Systems Research Institute (ESRI), Redlands, CA] and R. Materials and Methods Total Se concentrations in soils (mg Se/kg soil, reported herein as mg Se/kg; ACKNOWLEDGMENTS. We thank S. Adcock, M. Broadley, E. C. da Silva, Jr., soils were air dried or oven dried) 0–30 cm deep (n = 33,241 samples) were A. Chilimba, K. Dhillon, A. Donald, A. Eqani, L. Guilherme, E. Joy, K. Macey, obtained from Brazil, Canada, China, Europe, Japan, Kenya, Malawi, New A. Meharg, P. Morris, G. Paterson, H. Shen, J. van Ryssen, J. Wilford, and Zealand, South Africa, and the United States (see SI Materials and Methods J. Yanai for providing data; T. Blazina for digitizing the Chinese soil Se data; for dataset details and a discussion about which Se datasets were used, Fig. J. Hernandez for helping us obtain samples from the Rothamsted archive; S8). Samples derived from stream sediments were excluded from this analysis. M. Glendining for providing data for the Broadbalk soil experiment; In addition, we obtained 26 variables describing factors hypothesized to K. Abbaspour, K. Coleman, C. F. Randin, H. F. Satizábal, and R. Siber for input control soil Se concentrations and moderate climate change projections (RCP on methodological approaches; C. Stengel for assistance in analyzing soil samples for Se; R. Jones for comments on an earlier draft of this paper; 6.0 for climate and A1B for SOC data; see Table S1 forvariable descriptionsand U. Beyerle and J. Sedlacek for processing the Climate Model Intercomparison citations). All data within a 1° cell were averaged and represented by a single Project (CMIP) Phase 5 data; the World Climate Research Programme’s Work- value. To minimize the influence of errors and/or outliers within the datasets, ing Group on Coupled Modeling, which is responsible for CMIP; and the pixels containing fewer than five Se data points were removed from the climate modeling groups for producing and making available their model analysis (SI Materials and Methods). The final soil Se dataset consisted of n = output. The US Department of Energy’s Program for Climate Model Diagnosis 1,642 aggregated points. Four techniques for selecting variables [e.g., corre- and Intercomparison provides coordinating support for CMIP and led the de- lations, principal components analysis (PCA), backward elimination modeling, velopment of software infrastructure in partnership with the Global Organi- and RF node purity analyses; see SI Materials and Methods] were used to retain zation for Earth System Science Portals. This work was supported by Swiss National Science Foundation Grants PP00P2_133619 and PP00P2_163747 and the following variables for predictive analysis: AI, clay content, ET, lithology, Eawag, the Swiss Federal Institute of Aquatic Science and Technology. P.S. is pH, precipitation, and SOC at a soil depth of 0–30 cm. Although 16 lithological supported by the Delivering Food Security on Limited Land (DEVIL) project classes were present within the raster dataset, classes that were represented by (UK Natural Environmental Research Council NE/M021327/1) funded by the too few soil Se data points (n < 200) were grouped together instead of being Belmont Forum/Joint Programming Initiative on Agriculture, Food Security deleted (Fig. S4 and see SI Materials and Methods for further discussion). and Climate Change (FACCE-JPI) and by UK Biotechnology and Biological Sci- Predictive modeling was performed using three machine-learning models (one ences Research Council Project BB/L000113/1. Rothamsted Research is grant- RF and two artificial neural network models) (SI Materials and Methods). Each aided by the UK Biotechnology and Biological Sciences Research Council. 2852 | www.pnas.org/cgi/doi/10.1073/pnas.1611576114 Jones et al. 1. Stoltzfus RJ, Mullany L, Black RE (2004) Iron Deficiency Anaemia. Comparative 34. Trabucco A, Zomer R (2010) Global soil water balance geospatial database. CGIAR Quantification of Health Risks: Global and Regional Burden of Disease Attribution Consortium for Spatial Information. Published online, available from the CGIAR-CSI to Selected Major Risk Factors, eds Ezzati M, Lopez AD, Rodgers A, Murray CJL (World GeoPortal at www.cgiar-csi.org. Accessed January 28, 2017. Health Organization, Geneva), Vol 1, pp 163–210. 35. Siebert S, Henrich V, Frenken K, Burke J (2013) Update of the digital global map 2. Adamson P (2004) Vitamin and Mineral Deficiency: A Global Progress Report (Mi- of irrigation areas to version 5 (Rheinische Friedrich-Wilhelms-Universität, Bonn, cronutrient Initiative, UNICEF, Ottawa). Germany and Food and Agriculture Organization of the United Nations, Rome, 3. Fordyce FM (2013) Selenium deficiency and toxicity in the environment. Essentials of Italy). Medical Geology, eds Selinus O, et al. (Springer, London), pp 373–415. 36. Milly PC, Dunne KA, Vecchia AV (2005) Global pattern of trends in streamflow and 4. James L, et al. (1989) Selenium Poisoning in Livestock: A Review and Progress. water availability in a changing climate. Nature 438(7066):347–350. Selenium in Agriculture and the Environment (Soil Science Society of America and 37. Nachtergaele F, et al. (2010) Global land degradation information system (GLADIS), American Society of Agronomy, Madison, WI), pp 123–131. Beta version. (Food and Agriculture Organization of the United Nations, Rome), 5. Chilimba ADC, et al. (2011) Maize grain and soil surveys reveal suboptimal dietary Technical Report 17. selenium intake is widespread in Malawi. Sci Rep 1(72):72. 38. Hartmann J, Moosdorf N (2012) The new global lithological map database GLiM: A 6. Renkema H, et al. (2012) The effect of transpiration on selenium uptake and mobility representation of rock properties at the Earth surface. Geochem Geophys Geosyst in durum wheat and spring canola. Plant Soil 354(1-2):239–250. 13(12):1–37. 7. Winkel LHE, et al. (2015) Selenium cycling across soil-plant-atmosphere interfaces: A 39. Hollmann R, et al. (2013) The ESA climate change initiative: Satellite data records for critical review. Nutrients 7(6):4199–4239. essential climate variables. Bull Am Meteorol Soc 94(10):1541–1552. 8. Sun G-X, et al. (2016) Distribution of soil selenium in China is potentially controlled 40. Broxton PD, Zeng XB, Sulla-Menashe D, Troch PA (2014) A global land cover clima- bydeposition and volatilization? Sci Rep 6:20953. tology using MODIS data. J Appl Meteorol Climatol 53(6):1593–1605. 9. Winkel LHE, et al. (2012) Environmental selenium research: From microscopic pro- 41. Lefsky MA (2010) A global forest canopy height map from the Moderate Resolution cesses to global understanding. Environ Sci Technol 46(2):571–579. Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophys Res 10. Wen H, Carignan J (2007) Reviews on atmospheric selenium: Emissions, speciation and Lett 37(15):L15401. fate. Atmos Environ 41(34):7151–7165. 42. Center for International Earth Science Information Network - CIESIN – Columbia 11. Fernández-Martínez A, Charlet L (2009) Selenium environmental cycling and bio- University (2016) Documentation for the Gridded Population of the World, Version 4 availability: A structural chemist point of view. Rev Environ Sci Biotechnol 8(1):81–110. 12. Neal RH, Sposito G (1991) Selenium mobility in irrigated soil columns as affected by (GPWv4). (NASA Socioeconomic Data and Applications Center, Palisades NY) Avail- organic carbon amendment. J Environ Qual 20(4):808–814. able at sedac.ciesin.columbia.edu/data/collection/gpw-v4/documentation. Accessed 13. Blazina T, et al. (2014) Terrestrial selenium distribution in China is potentially linked January 28, 2017. to monsoonal climate. Nat Commun 5(4717):4717. 43. Hu L, Bentler PM (1999) Cutoff criteria for fit indexes in covariance structure analysis: 14. Lee GK, et al. (2016) The Geochemical Atlas of Alaska, 2016, US Geologic Survey Data Conventional criteria versus new alternatives. Struct Equ Modeling 6(1):1–55. Series 908, p. 25, GIS database available at https://pubs.er.usgs.gov/publication/ds908. 44. Rovira M, et al. (2008) Sorption of selenium(IV) and selenium(VI) onto natural iron Accessed January 28, 2017. oxides: Goethite and hematite. J Hazard Mater 150(2):279–284. 15. Fichtner SS, et al. (2007) Estudo da composição mineral de solos, forragens e tecido 45. Dhillon KS, Dhillon SK (2014) Development and mapping of seleniferous soils in animal de bovinos do município De Rio Verde, Goiás: Cobre, molibdênio e selênio. northwestern India. Chemosphere 99:56–63. Pesqui Agropecu Trop 20(1):1–6. 46. Western AW, Grayson RB, Blöschl G (2002) Scaling of soil moisture: A hydrologic 16. Faria L (2009) Levantamento Sobre Selênio em Solos e Plantas do Brasil e sua Apli- perspective. Annu Rev Earth Planet Sci 30(1):149–180. cação em Plantas Forrageiras. Thesis (Universidade De São Paulo, São Paulo, Brasil). 47. Silver WL, Lugo A, Keller M (1999) Soil oxygen availability and biogeochemistry along 17. Gabos MB, Alleoni LRF, Abreu CA (2014) Background levels of selenium in some se- rainfall and topographic gradients in upland wet tropical forest soils. Biogeochemistry lected Brazilian tropical soils. J Geochem Explor 145:35–39. 44(3):301–328. 18. Carvalho GS (2011) Selênio e Mercúrio em Solos sob Cerrado Nativo. Thesis (Uni- 48. Guo L, Frankenberger WT, Jury WA (1999) Evaluation of simultaneous reduction and versidade Federal de Lavras, Lavras, Brazil). transport of selenium in saturated soil columns. Water Resour Res 35(3):663–669. 19. Smith DB, et al. (2005) Major- and Trace-Element Concentrations in Soils from Two 49. Boulding JR, Ginn JS (2003) Practical Handbook of Soil, Vadose Zone, and Ground- Continental-Scale Transects of the United States and Canada, US Geologic Survey Water Contamination: Assessment, Prevention, and Remediation (CRC, Boca Raton, FL). Open-File Report 2005-1253. Available at https://pubs.usgs.gov/of/2005/1253/pdf/ 50. Fordyce F (2007) Selenium geochemistry and health. Ambio 36(1):94–97. OFR1253.pdf. Accessed January 28, 2017. 51. Ocean Biology Processing Group (2003) MODIS Aqua Level 3 Global Monthly Mapped 20. Friske P, et al. (2014) Soil Geochemical, Mineralogical, Radon and Gamma Ray Spec- 4 km Chlorophyll a. Ver. 6. (Physical Oceanography Distributed Active Archive Center, trometric Data from the 2007 North American Soil Geochemical Landscapes Project in Pasadena, CA). New Brunswick, Nova Scotia, and Prince Edward Island; Geological Survey of 446 52. Taylor KE, Stouffer RJ, Meehl GA (2012) An overview of Cmip5 and the experiment Canada, Open File 6433 (revised). Available at http://geogratis.gc.ca/api/en/nrcan- design. Bull Am Meteorol Soc 93(4):485–498. rncan/ess-sst/fa531017-b8bf-50c0-b7e4-a7764361415b.html. Accessed January 28, 2017. 53. Long JA, et al. (2015) Severe selenium depletion in the Phanerozoic oceans as a factor 21. Friske P, et al. (2013) Soil Geochemical, Radon and Gamma Ray Spectrometric Data in three global mass extinction events. Gondwana Res 36:209–218. from the 2008 and 2009 North American Soil Geochemical Landscapes Project Field 54. Yu T, et al. (2014) The origin and geochemical cycle of soil selenium in a Se-rich area Surveys, Open File 7334. Available at http://geoscan.nrcan.gc.ca/starweb/geoscan/ of China. J Geochem Explor 139:97–108. servlet.starweb?path=geoscan/fullf.web&search1=R=292514. Accessed January 28, 2017. 55. Intergovernmental Panel on Climate Change (2014) Climate Change 2014: Synthesis 22. Rawlins B, et al. (2012) The Advanced Soil Geochemical Atlas of England and Wales. (British Geological Survey, Keyworth, UK). Report, eds IPCC Core Writing Team, Pachauri RK, Meyer LA (IPCC, Geneva). 23. Reimann C, et al. (2014) Distribution of elements/parameters in agricultural and 56. Kirk GJD, Bellamy PH, Lark RM (2010) Changes in soil pH across England and Wales in grazing land soil in Europe. Chemistry of Europe’s Agricultural Soils. Part A: response to decreased acid deposition. Glob Change Biol 16(11):3111–3119. Methodology and Interpretation of the GEMAS Data Set, eds Reimann C, Birke M, 57. Jones DL, et al. (2013) Review: Nutrient stripping: The global disparity between food Demetriades A, Filzmoser P, O’Connor P (Bundesanstalt für Geowissenschaften und security and soil nutrient stocks. J Appl Ecol 50(4):851–862. Rohstoffe, Hannover, Germany), pp 103–474. 58. Myers SS, et al. (2014) Increasing CO threatens human nutrition. Nature 510(7503): 24. Yamada H, Kamada A, Usuki M, Yanai J (2009) Total selenium content of agricultural 139–142. soils in Japan. Soil Sci Plant Nutr 55(5):616–622. 59. Oldfield J (2002) Selenium World Atlas. (Selenium-Tellurium Development Associa- 25. Maskall JE, Thornton I (1989) The mineral status of Lake Nakuru National Park, Kenya - a tion, Grimbergen, Belgium). Updated edition. reconnaissance survey. Afr J Ecol 27(3):191–200. 60. Satizábal MH, Pérez-Uribe A (2007) Relevance metrics to reduce input dimensions in 26. Joy EJM, et al. (2015) Soil type influences crop mineral composition in Malawi. Sci artificial neural networks. Proceedings of the 17th International Conference on Ar- Total Environ 505:587–595. tificial Neural Networks, September 9–13, eds Sá JMd, Alexandre L, Duch W, Mandic D 27. Martin A, et al. (2016) The regional geochemical baseline soil survey of southern New (Springer, Berlin), pp 39-48. Zealand: Design and initial interpretation. J Geochem Explor 167:70–82. 61. Carpita M, Sandri M, Simonetto A, Zuccolotto P (2013) Football mining with R. Data 28. Zheng C (1994) Atlas of Soil Environmental Background Value in the People’s Mining Applications with R, eds Zhao Y, Cen Y (Academic, Waltham, MA), pp Republic of China (China Environmental Science Press, Beijing), p 195. 397–433. 29. ARC-ISCW (2005) Overview of the Agricultural Natural Resources of South Africa: Report 62. Nocedal J, Wright S (2006) Numerical Optimization (Springer Science & Business No. GW/A/2004/13. (ARC-Institute for Soil, Climate and Water, Pretoria, South Africa). Media, New York), 2nd Ed. 30. Greve P, Seneviratne SI (2015) Assessment of future changes in water availability and 63. Knights JS, Zhao FJ, Spiro B, McGrath SP (2000) Long-term effects of land use and aridity. Geophys Res Lett 42(13):5493–5499. fertilizer treatments on sulfur cycling. J Environ Qual 29(6):1867–1874. 31. Stackhouse PW, Jr, et al. (2006) Fast longwave and shortwave radiative flux 64. Hewitt A, Reynolds C (1990) Dissolution of metals from soils and sediments with a (FLASHFlux) products from CERES and MODIS measurements. Proceedings of the 12th microwave-nitric acid digestion technique. Spectrosc. 11(5):187–192. Conference on Atmospheric Radiation,July10–14, Madison, WI, (American Meteo- 65. Binstock DA, et al. (1990) Validation of Methods for Determining Elements in Solid rological Society, Boston), pp 1–6. Waste by Microwave Digestion. Waste Testing and Quality Assurance, ed Friedman D 32. Hengl T, et al. (2014) SoilGrids1km–global soil information based on automated (ASTM International, Philadelphia), Vol 2, pp 259–270. mapping. PLoS One 9(8):e105992. 66. Ramankutty N, Evan AT, Monfreda C, Foley JA (2008) Farming the planet: 1. Geo- 33. Gottschalk P, et al. (2012) How will organic carbon stocks in mineral soils evolve under graphic distribution of global agricultural lands in the year 2000. Global future climate? Global projections using RothC for a range of climate change sce- narios. Biogeosciences 9(8):3151–3171. Biogeochem Cycles 22(1):GB1003. Jones et al. PNAS | March 14, 2017 | vol. 114 | no. 11 | 2853 ENVIRONMENTAL SCIENCES

Journal

Proceedings of the National Academy of Sciences of the United States of AmericaPubmed Central

Published: Feb 21, 2017

There are no references for this article.