Data-mining analysis of the global distribution of soil carbon in observational databases and Earth system models

Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ doi:10.5194/gmd-10-1321-2017 © Au ho (s) 2017. CC A ibu ion 3.0 License. Da a-mining analysis o he global dis ibu ion o soil ca bon in obse a ional da abases and Ea h sys em models Shoji Hashimo o1, Kazuki Nanko1, Bo is ˇ Tupek2, and Aleksi Leh onen2 1Fo es y and Fo es P oduc s Resea ch Ins i u e (FFPRI), Tsukuba, Japan 2Na u al Resou ces Ins i u e Finland, La oka anonkaa i 9, Helsinki, Finland Co espondence o: Shoji Hashimo o (shojih@ p i.a c.go.jp) Recei ed: 27 May 2016 – Discussion s a ed: 30 June 2016 Re ised: 8 Feb ua y 2017 – Accep ed: 28 Feb ua y 2017 – Published: 28 Ma ch 2017 Abs ac . Fu u e clima e change will d ama ically change he ca bon balance in he soil, and his change will a ec he e es ial ca bon s ock and he clima e i sel . Ea h sys- em models (ESMs) a e used o unde s and he cu en cli- ma e and o p ojec u u e clima e condi ions, bu he soil o ganic ca bon (SOC) s ock simula ed by ESMs and hose o obse a ional da abases a e no well co ela ed when he wo a e compa ed a ine g id scales. Howe e , he speci ic key p ocesses and ac o s, as well as he ela ionships among hese ac o s ha go e n he SOC s ock, emain unclea ; he inclusion o such missing in o ma ion would imp o e he ag eemen be ween modeled and obse a ional da a. In his s udy, we sough o iden i y he in luen ial ac o s ha go - e n global SOC dis ibu ion in obse a ional da abases, as well as hose simula ed by ESMs. We used a da a-mining (machine-lea ning) (boos ed eg ession ees – BRT) scheme o iden i y he ac o s a ec ing he SOC s ock. We applied BRT scheme o h ee obse a ional da abases and 15 ESM ou pu s om he i h phase o he Coupled Model In e com- pa ison P ojec (CMIP5) and examined he e ec s o 13 a i- ables/ ac o s ca ego ized in o i e g oups (clima e, soil p op- e y, opog aphy, ege a ion, and land-use his o y). Globally, he con ibu ions o mean annual empe a u e, clay con en , ca bon- o-ni ogen (CN) a io, we land a io, and land co e we e high in obse a ional da abases, whe eas he con ibu- ions o he mean annual empe a u e, land co e , and ne p ima y p oduc i i y (NPP) we e p edominan in he SOC dis ibu ion in ESMs. A compa ison o he in luen ial ac- o s a a global scale e ealed ha he mos dis inc di e - ences be ween he SOCs om he obse a ional da abases and ESMs we e he low clay con en and CN a io con ibu- ions, and he high NPP con ibu ion in he ESMs. The esul s o his s udy will aid in iden i ying he causes o he cu en misma ches be ween obse a ional SOC da abases and ESM ou pu s and imp o e he modeling o e es ial ca bon dy- namics in ESMs. This s udy also e eals how a da a-mining algo i hm can be used o assess model ou pu s. 1 In oduc ion Soil is he la ges o ganic ca bon s ock in e es ial ecosys- ems (Ba jes, 1996; IPCC, 2013; Köchy e al., 2015). The soil o ganic ca bon (SOC) s ock ep esen s a balance be ween ca bon inpu s o soil and ca bon losses om soil ia decom- posi ion and dissol ed o ganic ca bon, and his in lux and e lux o soil ca bon is con olled di ec ly and indi ec ly by en i onmen al condi ions (Ca alhais e al., 2014; Schimel e al., 1994). Fu u e clima e change will d ama ically a ec he global soil ca bon balance (Bond-Lambe y and Thom- son, 2010; C ow he e al., 2016; F iedlings ein e al., 2006; Hashimo o e al., 2011, 2015), and his change will a ec e - es ial ca bon and consequen ly he clima e i sel (Cox e al., 2000; Zaehle, 2013). Ea h sys em models (ESMs) we e de eloped o unde - s and he cu en clima e and p o ide u u e clima e p ojec- ions, and hese models inco po a e he e es ial ca bon cy- cle, including SOC (A o a e al., 2013; F iedlings ein e al., 2014). In ecosys em ca bon cycle models o ESMs, SOC is calcula ed as he balance be ween ca bon inpu s ia dead o - ganic ma e and ca bon emissions ia o ganic ma e decom- posi ion, wi h bo h p ocesses in luenced by empe a u e and wa e condi ions. SOC dynamics ha e a c i ical in luence on Published by Cope nicus Publica ions on behal o he Eu opean Geosciences Union. 1322 S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon he land ca bon sink in ESM simula ions (F iedlings ein e al., 2014). Obse a ional global soil da abases a e o en used as benchma ks o examine whe he ESMs success ully desc ibe he global dis ibu ion o soil ca bon s ocks (Ana e al., 2013; Ha a uk e al., 2014; Todd-B own e al., 2013; Wiede e al., 2014). Se e al global soil da abases ha e been de- eloped o e he pas wo decades, and se e al a e un- de going u he imp o emen (Scha lemann e al., 2014). Ce ain da abases desc ibe he global dis ibu ion o soil physiochemical p ope ies and enable calcula ions o he global dis ibu ion o SOC s ocks (e.g., Ha monized Wo ld Soil Da abase – HWSD), whe eas o he s p o ide SOC s ocks by de aul (e.g., In e na ional Geosphe e Biosphe e P og amme’s (IGBP) Da a and In o ma ion Sys em (DIS) da abase). These da abases inco po a e obse ed da a poin s wi h global co e age, al hough he e a e biases in he spa ial dis ibu ion o densi y o he da a poin s. In hese da abases, g idded SOC da a ha e been gene a ed by linking he soil p ope ies o soil maps o by in e -ex apola ing he model ou pu s de i ed om analyses o obse ed SOC da a poin s. Compa ed wi h he SOC dis ibu ion de i ed om ESMs, SOC es ima es de i ed om SOC obse a ions a e mo e da a o ien ed; howe e , e en he obse a ional da abases include signi ican unce ain y because o e o s in he sou ce da a and building p ocesses (Köchy e al., 2015; Todd-B own e al., 2013). A ecen s udy (Todd-B own e al., 2013) ound ha al- hough ESM esul s a e mode a ely consis en a he biome le el, he co ela ion be ween he dis ibu ion o soil ca bon s ocks simula ed by ESMs and obse a ional da abases is poo when he wo a e compa ed a ine scales (e.g., a 1◦ scale). Fu he mo e, es ima es o SOC by ESMs and e es- ial biosphe e models exhibi high unce ain y (Nishina e al., 2014, 2015; Tian e al., 2015). Se e al s udies ha e ex- amined he cause o he inconsis ency in da a de i ed by obse a ional da abases and ESMs, and he high a ia ion o SOC ou pu s om ESMs (Exb aya e al., 2013; Todd- B own e al., 2013; Wiede e al., 2013). Todd-B own e al. (2013) analyzed he soil ca bon ou pu s om 11 ESMs om he i h phase o he Coupled Model In e compa ison P ojec (CMIP5) and soil ca bon da a om he HWSD, and ound ha ne p ima y p oduc i i y (NPP) and empe a u e could explain he SOC spa ial a ia ions in he ESM ou - pu bu no in he HWSD ou pu . These au ho s also ound ha he di e ences in SOC om he ESMs we e d i en by di e ences in he simula ed NPP and he pa ame e iza ion o soil he e o ophic espi a ion and no by di e ences in he soil model s uc u e o he ESMs. The key in luence o pa ame e izing soil he e o ophic espi a ion (e.g., u no e ime) on SOC in he CMIP5 ESM has also been discussed by Exb aya e al. (2013). Ana e al. (2013) examined he ela- ionships be ween simula ed SOC and ege a ion ca bon and compa ed hem wi h e e ence da a ob ained om obse a- ional da abases, and hey ound ha al hough simula ed al- ues clus e ed a ound he e e ence alues, he a io o SOC o ege a ion ca bon di e ed among he models. This ind- ing sugges s ha he pa ame e iza ion o plan p oduc ion, mo ali y, and decomposi ion a y g ea ly among ESMs. Mo e ealis ic ep esen a ions o u no e imes (Ko en e al., 2015) and ocusing on he model ea men o he hy- d ological cycle on he ca bon cycle (Shao e al., 2013) a e sugges ed as u u e imp o emen s. Despi e hese esea ch esul s, he key p ocesses and ac- o s ha go e n he SOC s ock and he ela ionships among hem emain unclea . The app op ia e inclusion o hese p o- cesses/ ac o s would imp o e he consis ency be ween he model esul s and obse a ional da a. In his s udy, we sough o iden i y he key ac o s ha go e n he global SOC dis- ibu ion in obse a ional da abases as well as hose simu- la ed by ESMs. We applied a da a-mining (machine-lea ning) scheme (boos ed eg ession ees – BRT) o iden i y he in- luen ial ac o s and explo e how hey ela e o SOC s ocks (Eli h e al., 2008). The BRT me hod is based on eg es- sion ees and boos ing. We combined he po en ially in- luen ial a iables om many da a p oduc s and SOC da a om obse a ional da abases and ESMs, and examined he ac o s in luencing he dis ibu ion o SOC and he ela ion- ships be ween hese ac o s and SOC s ocks. We assessed how closely ESMs could ma ch he in luen ial ac o s and hei ela ionships wi h ac o s ob ained om obse a ional da abases. By compa ing he in luen ial ac o s in he obse - a ional da abases wi h hose in he ESMs, we cla i ied he model–da a disc epancies and he a eas in which ESMs can be imp o ed. 2 Ma e ials and me hods 2.1 Obse a ional global SOC da abase We used SOC da a om wo global da abases and one no h- e n obse a ional da abase. The i s global da abase was he HWSD (FAO/IIASA/ISRIC/ISSCAS/JRC, 2012). The HWSD is a global da abase o soil physiochemical p ope - ies ha has been de eloped by he In e na ional Ins i u e o Applied Sys ems Analysis (IIASA) and he Food and Ag i- cul u e O ganiza ion o he Uni ed Na ions (FAO) in collab- o a ion wi h he In e na ional Soil Re e ence and In o ma- ion Cen e (ISRIC) Wo ld Soil In o ma ion, he Eu opean Commission Join Resea ch Cen e (JRC), and he Ins i u e o Soil Science, Chinese Academy o Sciences (ISSCAS). The da abase was cons uc ed by compiling he Eu opean Soil Da abase (ESDB), a 1 :1 million soil map o China, a - ious egional SOTER da abases (SOTWIS da abase), and a soil map o he wo ld om he FAO. We used an SOC s ock da abase ob ained wi h HWSD om he Join Resea ch Cen- e (JRC) (Hiede e and Köchy, 2011) (Fig. 1a). The second da abase included global g idded su aces o selec ed soil cha ac e is ics (IGBP-DIS) (Global Soil Da a Task G oup, Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 1323 60˚ S 0˚ 60˚ N (a) HWSD 60˚ S 0˚ 60˚ N (b) IGBP−DIS 180˚ 90˚ W 0˚ 90˚ E 180˚ 60˚ S 0˚ 60˚ N 0 20406080100 kg C m−2 (c) NCSCD Figu e 1. Soil ca bon s ock in he uppe 100 cm (kg C m−2) om he obse a ional da abases (HWSD, IGBP-DIS, and NCSCD). 2000) (Fig. 1b), which con ains g idded soil physiochemical p ope ies. The da abase has been de eloped by he Global Soil Da a Task G oup o he In e na ional Geosphe e Bio- sphe e P og amme’s (IGBP) Da a and In o ma ion Sys em (DIS), and he da abase was gene a ed by linking he pedon eco ds in he Global Pedon Da abase o he FAO/UNESCO digi al soil map o he wo ld. The hi d da abase was he No he n Ci cumpola Soil Ca bon Da abase, e sion 2 (NC- SCD) (Hugelius e al., 2013; Ta nocai e al., 2009) (Fig. 1c). This da abase is a spa ial da abase o SOC s ock o he no h- e n ci cumpola pe ma os egion. The soil map da a we e ob ained om di e en egions/coun ies (USA, Canada, Russia, e c.) and we e ha monized. The NCSCD we e based on 1778 pedon da a poin s. We used he HWSD and IGBP-DIS o analyze he global dis ibu ion o SOC s ocks; hen, we ex ac ed a da abase o no he n ci cumpola egions om he h ee abo e da abases and analyzed he SOC s ocks in he no he n egion. The e- la ionships among he da abases a e shown in Fig. S1 in he Supplemen . The SOC in he uppe 100 cm in each da abase was used. 2.2 Global SOC es ima ed using Ea h sys em models The global dis ibu ion o SOC s ocks es ima ed by ESMs was ob ained om CMIP5. We examined he esul s o 15 ESMs (Fig. 2; Table 2). When mo e han one esul was ob- ained by he same model amily (e.g., MIROC-ESM and MIROC-ESM-CHEM), we gene a ed an ensemble a e age da abase o each amily (e.g., a e age o MIROC-ESM and MIROC-ESM-CHEM): Todd-B own e al. (2013) showed h ough a hie a chical clus e analysis ha SOC dis ibu ions we e e y simila among ESMs om he same clima e cen- e . The mean alues om 1980–2004 we e calcula ed. The esul s o he his o ical and ensemble membe 1i1p1 we e used in his s udy. The no a ion “ 1i1p1” is an iden i ie o he model simula ion and is an ensemble membe ha is o en used o analyses (Chang e al., 2012; Di meye e al., 2013; Jiang e al., 2015; Kuma e al., 2014). The o e iews o SOC submodels in he ESMs ha e been p e iously desc ibed (Exb aya e al., 2014; Todd-B own e al., 2013, 2014) and a e also shown in Table 2. In gene al, each soil submodel consis ed o one o nine pools and inco po a ed he e ec s o empe a u e and mois u e. Some ESMs ha e li e ca bon pools; hese we e excluded om his s udy. A compa ison be- ween he mean o ESMs and global obse a ional da abases in a 1◦g id is shown in Fig. S2. 2.3 O he da abases We used i e g oups o a iables/ ac o s o examine hei e - ec s on global SOC: clima e, soil p ope y, opog aphy, eg- e a ion, and land-use his o y. De ailed da a sou ces o he da abases a e desc ibed in Table 1. The mean annual em- pe a u e and annual p ecipi a ion we e used as he clima e a iables, and he clay con en , ca bon- o-ni ogen (CN) a- io, and ex u e (Appendix Table A1) we e used as he soil a iables (0–30 cm). The compound opog aphic index, ele- a ion, slope, and we land a io we e used as he opog aphic indices. The CN a io was calcula ed by di iding he ca - bon densi y by he ni ogen densi y. The we land a io was calcula ed by di iding he numbe o we land g ids a 30 s by he o al g ids a 1◦. The lake, ese oi , and i e we e no quan i ied as we lands and we e excluded om he o- al g ids. The land co e ype (Appendix Table A2) and NPP we e adop ed as ege a ion indices, and he c opland a io and human app op ia ion o ne p ima y p oduc ion pe cen - age, which is a pe cen age o human consump ion o NPP o local NPP (Imho and Bounoua, 2006), we e used as he in- dices o land-use his o y. The a e age human app op ia ion o he NPP pe cen age was calcula ed a 1◦. His og ams o he a iables a e shown in Fig. S3. www.geosci-model-de .ne /10/1321/2017/ Geosci. Model De ., 10, 1321–1337, 2017 1324 S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 60˚ S 0˚ 60˚ N (a) BCC−ensemble (b) BNU−ESM (c) CanESM2 60˚ S 0˚ 60˚ N (d) CCSM4 (e) CESM1−ensemble ( ) CMCC−CESM 60˚ S 0˚ 60˚ N (g) GFDL−ESM2M (h) GISS−ensemble (i) HadGEM2−CC 60˚ S 0˚ 60˚ N (j) INM−CM4 (k) IPSL−ensemble (l) MIROC−ensemble 180˚ 90˚ W 0˚ 90˚ E 180˚ 60˚ S 0˚ 60˚ N (m) MPI−ensemble 180˚ 90˚ W 0˚ 90˚ E 180˚ 0 20406080100 kg C m−2 (n) MRI−ESM1 180˚ 90˚ W 0˚ 90˚ E 180˚ (o) No ESM1−ensemble Figu e 2. Soil ca bon s ocks (kg C m−2) om Ea h sys em models (CMIP5). The e m “ensemble” indica es he esul o an ensemble o amily membe s. 2.4 Da abase handling All global da abases, excep o he da abases wi h a spa- ial esolu ion o 1◦by de aul , including obse a ional and ESM model ou pu s, we e eg idded o a spa ial esolu ion o 1◦ o he analyses. Reg idding o da a in he Ne CDF o - ma was pe o med using he Clima e Da a Ope a o s (CDO) so wa e, e sion 1.6.9, p o ided by he Max Plank Ins i u e o Me eo ology (h ps://code.zmaw.de/p ojec s/cdo). A bi- linea in e pola ion, which is one o he mos widely used algo i hms, was used ( emapbil in CDO). 2.5 Boos ed eg ession ees (BRT) scheme To iden i y he in luen ial ac o s and hei ela ionships wi h SOC s ocks, BRT scheme was used in his s udy (Eli h e al., 2008). This echnique in ol es a da a-mining (machine- lea ning) algo i hm ha combines he ad an ages o a eg es- sion ee (decision ee) algo i hm and boos ing. Reg ession ees a e a classi ica ion algo i hm ha classi y da a h ough ecu si e bina y spli s, and boos ing is a machine-lea ning algo i hm ha gene a es many ough models and combines hem o imp o e hei p edic i e capabili y. The main ad an- ages o his me hod a e ha he BRT scheme can analyze di - e en ypes o a iables and in e ac ion e ec s among a i- ables, and is applicable o nonlinea ela ionships. In ecen yea s, he BRT echnique has been used o examine he dis- ibu ion o soil cha ac e is ics a a egional scale (Ae sen e al., 2011; Cools e al., 2014; Ma in e al., 2011). Majo ou - pu s om BRT analyses can iden i y he ollowing: (1) he ela i e impo ance (pe cen age o in luence o con ibu ion) o p edic o a iables (explana o y a iables) on he basis o he weigh ed and scaled numbe o imes a a iable is se- lec ed o spli ing (Eli h e al., 2008) and (2) he ela ion- ships among a iables and he explained a iable shown in pa ial dependence plo s. Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 1325 Table 1. Va iables used in he analyses and hei sou ces. Va iable Abb e ia ion Sou ce (da abase) O iginal esolu ion Re e ence Mean annual empe a u e1MAT ISLSCPII (CRU05) 1◦New e al. (2011) Mean annual p ecipi a ion1MAP ISLSCPII (CRU05) 1◦New e al. (2011) Clay con en (0–30 cm) Clay ISLSCPII 1◦Scholes and B own de Cols oun (2011) CN a io (0–30 cm)2CN a io ISLSCPII 1◦Scholes and B own de Cols oun (2011) Soil ex u e (0–30 cm) Tex u e ISLSCPII 1◦Scholes and B own de Cols oun (2011) Compound opog aphic index3CTI ISLSCPII 1◦Ve din (2011) Ele a ion3Ele . ISLSCPII 1◦Ve din (2011) Slope3Slope ISLSCPII 1◦Ve din (2011) We land a io We land Global Lakes and We lands Da abase 30 s Lehne and Döll (2004) Land co e LandCo e ISLSCPII 1◦F iedl e al. (2010) Ne p ima y p oduc ion NPP ISLSCPII 1◦P ince and Zheng (2011) C opland a io C opland ISLSCPII 1◦Ramanku y and Foley (2010) Human app op ia ion o NPP pe cen age HANPPpc HANPP collec ion 0.25◦Imho e al. (2004) 1The o iginal da abase p o ides mon hly da a. Annual means we e calcula ed by he au ho s. 2The CN a io was calcula ed by di iding he ca bon densi y by he ni ogen densi y. 3The na i e da abase is hyd o1k, and i s esolu ion is 1 km. The mean alue o 1 km was used in his s udy. We used he open-sou ce BRT package (b . unc ions.R) in R so wa e e sions 3.2.1 and 3.2.2 (R Co e eam, 2013) de eloped by Eli h e al. (2008). The R code o he BRT al- go i hm is a ailable in he supplemen a y ma e ial o Eli h e al. (2008). The gbm package was used ( e sion 2.1.1) o un he BRT package. The calcula ions we e pe o med in Mac OS X ( e sion 10.9.5 and e sion 10.10.5). To do so, he “windows” unc ion in he “b . unc ions.R” needed o be eplaced wi h he “qua z” unc ion in R. In p ac ice, h ee pa ame e s in he BRT package – he lea ning a e (l ), ee complexi y ( c), and bag ac ion (bg) – con ol he BRT pe - o mance. The l de e mines he con ibu ion o each ee, he c con ols he numbe o spli s, and he bg is he p opo - ion o da a selec ed a each s ep. The numbe o ees was de e mined using he c oss- alida ion me hod in he R pack- age. The maximum numbe o ees was se o 15 000. The c alue was se o 5. We es ed di e en l (0.001, 0.005, 0.01, 0.05, 0.1) and bg alues (0.5, 0.6, 0.7) and used he bes pa- ame e se o each da abase, bu he changes in pa ame e alues had li le e ec on he model pe o mance. 2.6 Model pe o mance The goodness o i be ween he BRT model and da a was assessed by using he linea ela ionship be ween he p e- dic ed and obse ed alues, he coe icien o de e mina ion (R2), and he oo mean squa e e o (RMSE); i is shown in Tables S2 and S3 in he Supplemen . Fo bo h he obse - a ional da abases and ESM da abases, he BRT models ex- hibi ed good pe o mance, wi h high R2 alues in mos o he da abases, bu he pe o mance was ela i ely lowe o NCSCD and CMCC (no he n soils). 3 Resul s 3.1 Obse a ional da abases 3.1.1 Global soil The ela i e con ibu ions o a iables in he BRT model o global SOC s ocks o he obse a ional da abases a e shown in Fig. 3a and b. In HWSD, he con ibu ions o land co e , mean annual empe a u e, CN a io, and we land a io we e high. Fo IGBP-DIS, he mean annual empe a u e, ollowed by clay con en , CN a io, and land co e also highly con- ibu ed. In pa icula , he mean annual empe a u e was e y in luen ial. The con ibu ion o ele a ion o each HWSD and IGBP-DIS was 6 and 7 %, espec i ely. The NPP con ibu ed 5 % in bo h da abases. The ela ionships be ween he in luen ial a iables and SOC a e shown in Fig. 4a–e. In gene al, he wo da abases showed simila ela ionships. Fo example, he SOC de- c eased wi h inc easing mean annual empe a u e, pa ic- ula ly a si es wi h a mean annual empe a u e >0◦C (Fig. 4a), bu inc eased wi h inc easing clay con en and CN www.geosci-model-de .ne /10/1321/2017/ Geosci. Model De ., 10, 1321–1337, 2017 1326 S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 0 10 20 30 40 50 (a) Global, HWSD Con ibu ion (%) 0 10 20 30 40 50 MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc (b) Global, IGBP-DIS Con ibu ion (%) Va iables 0 10 20 30 40 50 (c) No h, HWSD Con ibu ion (%) 0 10 20 30 40 50 (d) No h, IGBP-DIS Con ibu ion (%) 0 10 20 30 40 50 MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc (e) No h, NCSCD Con ibu ion (%) Va iables Figu e 3. Rela i e con ibu ion (in luence) o p edic i e a iables o he model o soil ca bon s ocks in he global obse a ional da abases (le ) and no he n obse a ional da abases ( igh ). a io (Fig. 4b and c). The SOC inc eased apidly wi h an in- c easing CN a io. Rela ionships wi h he mean annual em- pe a u e we e simila (Fig. 4a). The ela ionship wi h clay was s eepe in IGBP-DIS han in HWSD, bu he opposi e was ue o he CN a io (Fig. 4b and c). Wi h espec o land co e , e e g een needlelea o es s and pe manen we lands had highe SOC (Fig. 4e). 3.1.2 No he n soils In he no he n egion, he dominan con ibu o s di e ed among no he n soil da abases and om hose iden i ied in he global da abase analyses desc ibed abo e (Fig. 3c–e). In HWSD, he CN a io was he dominan con ibu o , ollowed by he we land a io, clay con en , and mean annual p ecip- i a ion. In IGBP-DIS, clay con en , CN a io, and ele a ion we e he mos impo an con ibu o s. Fo NCSCD, ele a ion con ibu ed he mos (∼25 %), bu all o he a iables excep o he c opland a io and HANPPpc con ibu ed 5–15 %. The mean annual empe a u e was no as in luen ial as he global da abases. The ela ionships be ween a iables and SOC s ock a ied mo e among he da abases o no he n soils han hose o global da abases (Fig. 4 –k). Fu he mo e, because he no h- e n egions we e ex ac ed, he anges o a iables we e na - owe han he global da abases. In NCSCD, he SOC de- c eased wi h inc easing empe a u e (Fig. 4 ) and inc eased wi h inc easing p ecipi a ion (Fig. 4g). The SOC inc eased wi h inc easing clay con en and CN a io in HWSD and IGBP-DIS (Fig. 4h and i), which was consis en wi h he indings ob ained om he global da abases. The inc easing end wi h inc easing CN a io was also obse ed in NC- SCD. The SOC dec eased wi h inc easing ele a ion in all da abases bu showed conside able a iabili y a low ele a- ions (Fig. 4j). 3.2 Ea h sys em models 3.2.1 Global soil The con ibu ions o some a iables a ied among ESMs, bu he mean o he esul s o he ESMs showed ha he mean annual empe a u e, land co e , and NPP clea ly con ibu ed o SOC dis ibu ion (Fig. 5a and b). La ge inconsis encies be ween he obse a ional da abases and ESMs we e ound in he low con ibu ions o clay con en and he CN a io and in he high con ibu ions o NPP in ESMs (Fig. 5a and b). The con ibu ion o NPP o ESMs was g ea e han in he obse a ional da abases. The ela ionships be ween SOC and ce ain a iables sub- s an ially a ied among he ESM da abases (Fig. 6a–e), pa - icula ly in he mean annual empe a u e (Fig. 6a). The SOC dec eased wi h inc easing mean annual empe a u e (Fig. 6a) bu inc eased wi h inc easing p ecipi a ion (Fig. 6b) and NPP (Fig. 6e). The mean o he ela ionship wi h mean annual empe a u e o ESMs was highly consis en wi h ha in he Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 1327 -10 0 10 -20 -10 0 10 20 30 (a) Fi ed unc ion (kg C m− 2 ) MAT ( C) -10 0 10 0 10 20 30 40 50 (b) Clay (%) -20 -10 0 10 20 0 10 20 30 40 50 (c) C / N a io -10 0 10 0 0.5 1 (d) Fi ed unc ion (kg C m− 2 ) We land a io -10 0 10 0 5 10 15 (e) Land co e HWSD IGBP-DIS -10 0 10 -20 -10 0 10 20 30 ( ) Fi ed unc ion (kg C m− 2 ) MAT ( C) -10 0 10 0 1000 2000 3000 4000 (g) MAP (mm) -20 -10 0 10 20 0 10 20 30 40 50 (h) Clay (%) -20 -10 0 10 20 0 10 20 30 40 50 (i) Fi ed unc ion (kg C m− 2 ) C / N a io -10 0 10 0 1000 2000 3000 (j) Ele (m) -10 0 10 0 0.5 1 (k) We land a io HWSD IGBP-DIS NCSCD o o Figu e 4. E ec s o he mos in luen ial a iables in he model o he soil ca bon s ock o each global (a–e) and no he n ( –k) obse a ional da abases. The i ed unc ions we e cen e ed by sub ac ing hei means. See Table A2 o land co e classi ica ions. Because o he small numbe o da a poin s, he esul s o “pe manen snow and ice” a e no shown (e). The y-axis scales o clay and he CN a io a e di e en om hose o o he ac o s (c, h, i). HWSD and IGBP-DIS da abases o he empe a u e ange −5 o 15 ◦C (Fig. 6a). The inc easing end wi h inc easing NPP in ESMs was consis en wi h ha o he HWSD, pa - icula ly below app oxima ely 500 g C m−2o NPP (Fig. 6e). Al hough he we land a io did no con ibu e o he ESMs (Fig. 6a) wi h espec o land co e , pe manen we lands had highe SOC (Fig. 6d). 3.2.2 No he n soils The mean o he ESMs showed ha o no he n soils, he main con ibu o s (mean annual empe a u e, land co e , and NPP) we e mainly he same as in he ESM global ou pu s (Fig. 5c and d). The con ibu ion o he mean annual empe - a u e was lowe han ha o he global esul s o he ESMs (mean o 14 % o he no he n and 29 % o he global em- pe a u es). The ela i ely la ge disc epancy be ween he ob- se a ional da abases and ESMs included he lowe con ibu- ion o clay con en , CN a io, and ele a ion, and he highe con ibu ion o he mean annual empe a u e, land co e , and NPP in he ESMs. The ela ionship be ween SOC and a iables in ESMs as well as he esul s o he obse a ional da abases a e shown in Fig. 6 –i. The mean o he ESMs indica ed ha he SOC in he no he n egion inc eased wi h inc easing NPP, and he ela ionship was simila o ha in HWSD (Fig. 6i), al- hough he con ibu ion o NPP in he ESMs di e ed om www.geosci-model-de .ne /10/1321/2017/ Geosci. Model De ., 10, 1321–1337, 2017 1328 S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon (b) (d) % % 0 10 20 30 40 50 60 70 80 MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc (a) Global Con ibu ion (%) ESM mean HWSD IGBP-DIS 0 10 20 30 40 50 60 70 80 MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc (c) No h Con ibu ion (%) Va iables ESM mean HWSD IGBP-DIS NCSCD Da abases No ESM1−ensemble MRI−ESM1 MPI−ensemble MIROC−ensemble IPSL−ensemble INM−CM4 HadGEM2−CC GISS−ensemble GFDL−ESM2M CMCC−CESM CESM1−ensemble CCSM4 CanESM2 BNU−ESM BCC−ensemble ESM−mean IGBP−DIS HWSD MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc 0 20 40 60 80 100 Va iables Da abases No ESM1−ensemble MRI−ESM1 MPI−ensemble MIROC−ensemble IPSL−ensemble INM−CM4 HadGEM2−CC GISS−ensemble GFDL−ESM2M CMCC−CESM CESM1−ensemble CCSM4 CanESM2 BNU−ESM BCC−ensemble ESM−mean NCSCD IGBP−DIS HWSD MAT MAP Clay CN a io Tex u e CTI Ele Slope We land LandCo e NPP C opland HANPPpc 0 20 40 60 80 100 Figu e 5. Rela i e con ibu ion (in luence) o p edic i e a iables o he model o he soil ca bon s ock om ESMs and a compa ison wi h hose o obse a ional da abases. Box plo s show he esul s o ESMs, and he pu ple, g een, ligh blue, and blue ma ks indica e he mean o he ESMs and esul s om obse a ional da abases (a: global; c: no h). Mosaic plo s o de ailed ela i e con ibu ions o each ESM (b: global; d: no h) a e shown. hose o he obse a ional da abase (Fig. 5c). The dec easing end wi h ele a ion was no eplica ed in he ESMs (Fig. 6g). 4 Discussion and concluding ema ks 4.1 Iden i ied in luen ial ac o s Compa ed wi h p e ious s udies, we examined he con ibu- ions o a wide a ie y o ac o s o SOC dis ibu ions. Ou analyses e ealed ha he mos dis inc di e ences be ween he obse a ional da abase da a and he ESM ou pu s we e he e ec s o he CN a io and clay con en (Fig. 5). Fo bo h global obse a ional da abases, he CN a io was a subs an ial con ibu o (Fig. 3a and b). The impo an con ibu ion o he CN a io was he same in he no he n da abases (Fig. 3c–e). The SOC in he obse a ional da abases inc eased wi h in- c eases in he CN a io (Fig. 4c), whe eas he SOC alues o he ESMs we e insensi i e o he CN a io. Ou esul s sup- po he impo ance o p ope ly inco po a ing he ni ogen (N) cycle in o SOC models (e.g., con ol o e decomposi- ion, soil e ili y, nu ien a ailabili y, and plan li e quali y) (Be g e al., 2001; Co u o e al., 2013; Fe nández-Ma ínez e al., 2014; Liski e al., 2005; Tuomi e al., 2009; ˇ Tupek e al., 2016). None o he ESMs excep o he CESM1 and No ESM in CMIP5 included e es ial ni ogen p ocesses (Todd-B own e al., 2013); howe e , including his pa ame e has been sugges ed as a key imp o emen o he nex model in e compa ison (CMIP6) (Hajima e al., 2014; Zaehle e al., 2015). The esul s o ou analysis suppo he impo ance o including he N cycle in ESM models. Clay con en is also o en used as a egula o o he decom- posabili y o o ganic ma e in he soil (e.g., CENTURY and Ro hC) (Coleman and Jenkinson, 1999; Pa on e al., 1987). Gene ally, high clay con en inhibi s o ganic ma e decom- posi ion in he soil. Fu he mo e, high clay con en s o en esul in low d ainage and anae obic soil condi ions, which also inhibi o ganic ma e decomposi ion. Fo he IGBP- DIS da a, he con ibu ion o he clay con en was as high as ha o he CN a io. The con ol o decomposabili y by he Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 1329 -10 0 10 -20 -10 0 10 20 30 (a) Fi ed unc ion (kg C m− 2 ) MAT ( C) -10 0 10 0 1000 2000 3000 4000 (b) MAP (mm) -10 0 10 0 1000 2000 3000 (c) Ele (m) -10 0 10 0 5 10 15 (d) Fi ed unc ion (kg C m− 2 ) Land co e -10 0 10 0 500 1000 1500 (e) NPP (g C m−2) ESM mean HWSD IGBP-DIS -10 0 10 -20 -10 0 10 20 30 ( ) Fi ed unc ion (kg C m− 2 ) MAT ( C) -10 0 10 0 1000 2000 3000 (g) Ele (m) -10 0 10 0 5 10 15 (h) Land co e -10 0 10 0 500 1000 1500 (i) Fi ed unc ion (kg C m− 2 ) NPP (g C m−2) ESM mean HWSD IGBP-DIS NCSCD o o Figu e 6. E ec o he mos in luen ial a iables in he model o global (a–e) and no he n ( –i) ou pu s om ESMs and a compa ison wi h hose o obse a ional da abases. G ey lines show he esul s o each ESM, and he pu ple line indica es he mean o he ESMs. The i ed unc ions we e cen e ed by sub ac ing hei means. See Table A2 o land co e classi ica ions. Because o he small numbe o da a poin s, he esul s o “pe manen snow and ice” a e no shown (d, h). clay con en has been p e iously inco po a ed in si e-scale p ocess-based models (Pa on e al., 1987) and may be in- co po a ed in ce ain ESMs because he soil ca bon submod- els in hese ESMs a e based on he CENTURY model (see he soil model his o y epo ed in Todd-B own e al., 2014). Howe e , ega dless o whe he he con ol o decomposabil- i y by clay is inco po a ed, ou esul s sugges ha he in lu- ence o clay on he ca bon cycle is no well cap u ed in mos ESMs. The mean annual empe a u e was iden i ied as an in luen- ial ac o in he global da abases (Fig. 3a and b) bu no in he no he n soil da abases (Fig. 3c–e). Tempe a u e is a main ac o con olling bo h plan p oduc ion (sou ce o ca bon in- pu o soil) and soil o ganic ma e decomposi ion, which a e al eady inco po a ed in ESMs. Based on an analysis o he ou pu o he e o ophic espi a ion, he empe a u e sensi i - i y (e.g., Q10 alue) o soil o ganic ma e decomposi ion in he ESMs has been epo ed as 1.4 o 2.2 (Todd-B own e al., 2014). In addi ion, ou analyses iden i ied di e se ela- ionships be ween he mean annual empe a u e and SOC. The lowe con ibu ion o he mean annual empe a u e in he no he n soils likely occu ed because empe a u e sensi i - i y is an exponen ial p ocess, and he magni ude o obse ed changes unde changing empe a u e is ela i ely small a a low empe a u e ange, as shown in a compa ison o empe - a u e unc ions in common biogeochemical models (Sie a e al., 2015). The ela ionships be ween he SOC and em- pe a u e ob ained in his s udy include he in eg a ion o he empe a u e sensi i i y o bo h plan p oduc ion and soil o - ganic decomposi ion and hus do no p o ide he empe a- u e sensi i i y pa ame e o indi idual p ocesses o ESMs. Howe e , he esul s o his s udy can be used o examine www.geosci-model-de .ne /10/1321/2017/ Geosci. Model De ., 10, 1321–1337, 2017 1336 S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon Liski, J., Palosuo, T., Pel oniemi, M., and Sie änen, R.: Ca bon and decomposi ion model Yasso o o es soils, Ecol. Model., 189, 168–182, doi:10.1016/j.ecolmodel.2005.03.005, 2005. Luo, Y., Ahls öm, A., Allison, S. D., Ba jes, N. H., B o kin, V., Ca alhais, N., Chappell, A., Ciais, P., Da idson, E. A., Finzi, A., Geo giou, K., Guene , B., Ha a uk, O., Ha den, J. W., He, Y., Hopkins, F., Jiang, L., Ko en, C., Jackson, R. B., Jones, C. D., La a, M. J., Liang, J., McGui e, A. D., Pa on, W., Peng, C., Rande son, J. T., Salaza , A., Sie a, C. A., Smi h, M. J., Tian, H., Todd-B own, K. E. O., To n, M., an G oenigen, K. J., Wang, Y. P., Wes , T. O., Wei, Y., Wiede , W. R., Xia, J., Xu, X., Xu, X., and Zhou, T.: Towa d mo e ealis ic p ojec ions o soil ca bon dynamics by Ea h sys em models, Global Biogeochem. Cy., 30, 40–56, doi:10.1002/2015GB005239, 2016. Manzoni, S. and Po po a o, A.: Soil ca bon and ni ogen mine al- iza ion: Theo y and models ac oss scales, Soil Biol. Biochem., 41, 1355–1379, doi:10.1016/j.soilbio.2009.02.031, 2009. Ma in, M. P., Wa enbach, M., Smi h, P., Mee smans, J., Joli e , C., Boulonne, L., and A ouays, D.: Spa ial dis ibu ion o soil o ganic ca bon s ocks in F ance, Biogeosciences, 8, 1053–1065, doi:10.5194/bg-8-1053-2011, 2011. New, M., Jones, P. D., and Hulme, M.: ISLSCP II Cli- ma e Resea ch Uni CRU05 Mon hly Clima e Da a, doi:10.3334/ORNLDAAC/1015, 2011. Nishina, K., I o, A., Bee ling, D. J., Cadule, P., Ciais, P., Cla k, D. B., Falloon, P., F iend, A. D., Kahana, R., Ka o, E., Ke ibin, R., Luch , W., Lomas, M., Rademache , T. T., Pa lick, R., Schapho , S., Vuicha d, N., Wa szawaski, L., and Yokoha a, T.: Quan i ying unce ain ies in soil ca bon esponses o changes in global mean empe a u e and p ecipi a ion, Ea h Sys . Dynam., 5, 197–209, doi:10.5194/esd-5-197-2014, 2014. Nishina, K., I o, A., Falloon, P., F iend, A. D., Bee ling, D. J., Ciais, P., Cla k, D. B., Kahana, R., Ka o, E., Luch , W., Lomas, M., Pa lick, R., Schapho , S., Wa szawaski, L., and Yokoha a, T.: Decomposing unce ain ies in he u u e e es ial ca bon bud- ge associa ed wi h emission scena ios, clima e p ojec ions, and ecosys em simula ions using he ISI-MIP esul s, Ea h Sys . Dy- nam., 6, 435–445, doi:10.5194/esd-6-435-2015, 2015. Os le, N. J., Smi h, P., Fishe , R., Woodwa d, F. I., Fishe , J. B., Smi h, J. U., Galb ai h, D., Le y, P., Mei , P., McNama a, N. P., and Ba dge , R. D.: In eg a ing plan -soil in e ac ions in o global ca bon cycle models, J. Ecol., 97, 851–863, 2009. Pa on, W. J., Schimel, D. S., Cole, C. V., and Ojima, D. S.: Analysis o ac o s con olling soil o ganic ma e le els in g ea plains g asslands, Soil Sci. Soc. Am. J., 51, 1173–1179, 1987. P ince, S. D. and Zheng, D. L.: ISLSCP II global p i- ma y p oduc ion da a ini ia i e g idded NPP da a, doi:10.3334/ORNLDAAC/1023, 2011. Ramanku y, N. and Foley, J. A.: ISLSCP II his o ical c oplands co e , 1700–1992, doi:10.3334/ORNLDAAC/966, 2010. R Co e eam: R: A language and en i onmen o s a is ical com- pu ing, R Founda ion o S a is ical Compu ing, Vienna, 2013. Scha lemann, J. P., Tanne , E. V., Hiede e , R., and Ka- pos, V.: Global soil ca bon: unde s anding and managing he la ges e es ial ca bon pool, Cabon Manag., 5, 81–91, doi:10.4155/cm .13.77, 2014. Schimel, D. S., B aswell, B. H., Holland, E. a., McKeown, R., Ojima, D. S., Pain e , T. H., Pa on, W. J., and Townsend, A. R.: Clima ic, edaphic, and bio ic con ols o e s o age and u no e o ca bon in soils, Global Biogeochem. Cy., 8, 279–293, doi:10.1029/94GB00993, 1994. Scholes, E. and B own de Cols oun, E.: ISLSCP II global g idded soil cha ac e is ics, doi:10.3334/ORNLDAAC/1004, 2011. Shao, P., Zeng, X., Sakaguchi, K., Monson, R. K., and Zeng, X.: Te es ial ca bon cycle: Clima e ela ions in eigh CMIP5 ea h sys em models, J. Clima e, 26, 8744–8764, doi:10.1175/JCLI-D- 12-00831.1, 2013. Sie a, C. A. and Mülle , M.: A gene al ma hema ical amewo k o ep esen ing soil o ganic ma e dynamics, Ecol. Monog ., 85, 505–524, doi:10.1890/15-0361.1, 2015. Sie a, C. A., T umbo e, S. E., Da idson, E. A., Vicca, S., and Janssens, I.: Sensi i i y o decomposi ion a es o soil o - ganic ma e wi h espec o simul aneous changes in empe - a u e and mois u e, J. Ad . Model. Ea h Sys ., 7, 335–356, doi:10.1002/2014MS000358, 2015. Ta nocai, C., Canadell, J. G., Schuu , E. A. G., Kuh y, P., Mazhi- o a, G. and Zimo , S.: Soil o ganic ca bon pools in he no h- e n ci cumpola pe ma os egion, Global Biogeochem. Cy., 23, GB2023, doi:10.1029/2008GB003327, 2009. Tian, H., Lu, C., Yang, J., Bange , K., Hun zinge , D. N., Schwalm, C. R., Michalak, A. M., Cook, R., Ciais, P., Hayes, D., Huang, M., I o, A., Jain, A. K., Lei, H., Mao, J., Pan, S., Pos , W. M., Peng, S., Poul e , B., Ren, W., Ricciu o, D., Schae e , K., Shi, X., Tao, B., Wang, W., Wei, Y., Yang, Q., Zhang, B., and Zeng, N.: Global pa e ns and con ols o soil o ganic ca bon dynamics as simula ed by mul iple e es ial biosphe e models: Cu en s a us and u u e di ec ions, Global Biogeochem. Cy., 29, 775– 792, doi:10.1002/2014GB005021, 2015. Todd-B own, K. E. O., Rande son, J. T., Hopkins, F., A o a, V., Ha- jima, T., Jones, C., She liako a, E., Tjipu a, J., Volodin, E., Wu, T., Zhang, Q., and Allison, S. D.: Changes in soil o ganic ca bon s o age p edic ed by Ea h sys em models du ing he 21s cen- u y, Biogeosciences, 11, 2341–2356, doi:10.5194/bg-11-2341- 2014, 2014. Todd-B own, K. E. O., Rande son, J. T., Pos , W. M., Ho man, F. M., Ta nocai, C., Schuu , E. A. G., and Allison, S. D.: Causes o a ia ion in soil ca bon simula ions om CMIP5 Ea h sys em models and compa ison wi h obse a ions, Biogeosciences, 10, 1717–1736, doi:10.5194/bg-10-1717-2013, 2013. Tuomi, M., Thum, T., Jä inen, H., F onzek, S., Be g, B., Ha mon, M., T o ymow, J. A., Se an o, S., and Liski, J.: Lea li e decomposi ion – Es ima es o global a iabili y based on Yasso07 model, Ecol. Model., 220, 3362–3371, doi:10.1016/j.ecolmodel.2009.05.016, 2009. ˇ Tupek, B., O iz, C. A., Hashimo o, S., S endahl, J., Dahlg en, J., Ka l un, E., and Leh onen, A.: Unde es ima ion o bo eal soil ca bon s ocks by ma hema ical soil ca bon models linked o soil nu ien s a us, Biogeosciences, 13, 4439–4459, doi:10.5194/bg- 13-4439-2016, 2016. Ve din, K. L.: ISLSCP II HYDRO1k Ele a ion-de i ed P oduc s, doi:10.3334/ORNLDAAC/1007, 2011. Wang, Y. P., Chen, B. C., Wiede , W. R., Lei e, M., Medlyn, B. E., Rasmussen, M., Smi h, M. J., Agus o, F. B., Ho man, F., and Luo, Y. Q.: Oscilla o y beha io o wo nonlinea mic obial models o soil ca bon decomposi ion, Biogeosciences, 11, 1817– 1831, doi:10.5194/bg-11-1817-2014, 2014. Wang, Y. P., Jiang, J., Chen-Cha pen ie , B., Agus o, F. B., Has ings, A., Ho man, F., Rasmussen, M., Smi h, M. J., Todd-B own, K., Geosci. Model De ., 10, 1321–1337, 2017 www.geosci-model-de .ne /10/1321/2017/ S. Hashimo o e al.: Da a-mining analysis o he global dis ibu ion o soil ca bon 1337 Wang, Y., Xu, X., and Luo, Y. Q.: Responses o wo nonlinea mic obial models o wa ming and inc eased ca bon inpu , Bio- geosciences, 13, 887–902, doi:10.5194/bg-13-887-2016, 2016. Wiede , W. R., Allison, S. D., Da idson, E. A., Geo giou, K., Ha a uk, O., He, Y., Hopkins, F., Luo, Y., Smi h, M. J., Sulman, B., Todd-B own, K., Wang, Y.-P., Xia, J., and Xu, X.: Explici ly ep esen ing soil mic obial p ocesses in Ea h sys em models, Global Biogeochem. Cy., 29, 1782–1800, doi:10.1002/2015GB005188, 2015. Wiede , W. R., Bonan, G. B., and Allison, S. D.: Global soil ca bon p ojec ions a e imp o ed by modelling mic obial p ocesses, Na . Clim. Change, 3, 909–912, doi:10.1038/nclima e1951, 2013. Wiede , W. R., Boehne , J., and Bonan, G. B.: E alua ing soil biogeochemis y pa ame e iza ions in Ea h sys em mod- els wi h obse a ions, Global Biogeochem. Cy., 28, 211–222, doi:10.1002/2013GB004665, 2014. Wu zle , T. and Reichs ein, M.: Soils apa om equilib ium – con- sequences o soil ca bon balance modelling, Biogeosciences, 4, 125–136, doi:10.5194/bg-4-125-2007, 2007. Zaehle, S.: Te es ial ni ogen-ca bon cycle in e ac ions a he global scale, Philos. T. Roy. Soc. B, 368, 20130125, doi:10.1098/ s b.2013.0125, 2013. Zaehle, S., Jones, C. D., Houl on, B., Lama que, J.-F., and Robe - son, E.: Ni ogen a ailabili y educes CMIP5 p ojec ions o wen y- i s -cen u y land ca bon up ake, J. Clima e, 28, 2494– 2511, doi:10.1175/JCLI-D-13-00776.1, 2015. www.geosci-model-de .ne /10/1321/2017/ Geosci. Model De ., 10, 1321–1337, 2017