Long‐Term Wetting and Drying Trends in Land Water Storage Derived From GRACE and CMIP5 Models

Long-Te m We ing and D ying T ends in Land Wa e S o age De i ed F om GRACE and CMIP5 Models L. Jensen1, A. Eicke 1, H. Dobslaw2, T. S acke3, and V. Humph ey4 1Geodesy and Geoin o ma ics, Ha enCi y Uni e si y Hambu g, Hambu g, Ge many, 2Helmhol z Cen e Po sdam, Ge man Resea ch Cen e o Geosciences (GFZ), Po sdam, Ge many, 3Max Planck Ins i u e o Me eo ology, Hambu g, Ge many, 4Ins i u e o A mosphe ic and Clima e Science, ETH Zü ich, Zü ich, Swi ze land Abs ac Coupled clima e models pa icipa ing in he CMIP5 (Coupled Model In e compa ison P ojec Phase 5) exhibi a la ge in e model sp ead in he ep esen a ion o long- e m ends in soil mois u e and snow in esponse o an h opogenic clima e change. We e alua e long- e m (Janua y 1861 o Decembe 2099) wa e s o age ends om 21 CMIP5 models agains obse ed ends in e es ial wa e s o age (TWS) ob ained om 14 yea s (Ap il 2002 o Augus 2016) o he GRACE (G a i y Reco e y And Clima e Expe imen ) sa elli e mission. This is complica ed due o he incomple e ep esen a ion o TWS in CMIP5 models and in e annual clima e a iabili y masking long- e m ends in obse a ions. We hus e alua e i s he sp ead in p ojec ed ends among CMIP5 models and iden i y egions o b oad model consensus. Second, we assess he ex en o which hese p ojec ed ends a e al eady p esen du ing he his o ical pe iod (Janua y 1861 o Augus 2016) and hus po en ially de ec able in obse a ional eco ds a ailable oday. Thi d, we quan i y he deg ee o which 14-yea endencies can be expec ed o ep esen long- e m ends, inding ha egional long- e m ends s a o eme ge om in e annual a ia ions a e jus 14 yea s while s able global end pa e ns a e de ec able a e 30 yea s. We classi y egions o s ong model consensus in o a eas whe e (1) clima e- ela ed TWS changes a e suppo ed by he di ec ion o GRACE ends, (2) misma ch o ends hin s a possible model de ici s, (3) he sho obse a ion ime span and/o an h opogenic in luences p e en eliable conclusions abou long- e m we ing o d ying. We he eby demons a e he alue o sa elli e obse a ions o wa e s o age o u he cons ain he esponse o he e es ial wa e cycle o clima e change. 1. In oduc ion The e es ial b anch o he global wa e cycle is an impo an componen o he Ea h's coupled clima e sys em: Wa e a ailable in he soil c i ically de e mines biomass p oduc ion ha e ec i ely akes up ca bon dioxide om he a mosphe e and hus cons i u es he land co e and consequen ly also he albedo o he Ea h's su ace. The a ailabili y o wa e a he su ace in luences he a e o e apo anspi a ion and he eby he amoun o la en hea abso bed by he a mosphe e locally and ad ec ed o dis an egions along wi h he oposphe ic winds; and wa e in he o m o snow co e he mally isola es he soil om he ai abo e i . The accu a e ep esen a ion o he e es ial wa e dynamics and i s a ious eedbacks o he a mosphe ic wa e , ene gy and ca bon cycles is hus c i ically impo an o in e ac i ely coupled global nume ical cli- ma e models ha a e used o in e in o ma ion abou he cu en s a e and he u u e e olu ion o he Ea h's clima e condi ions (T enbe h, 2010). Due o hei di ec e ec on he a ailabili y o eshwa e esou ces, in es iga ing clima e change impac s on he global wa e cycle is o g ea socie al ele ance. Changes in e es ial wa e s o age (TWS) migh e lec long- e m we ing o d ying in a ious egions o he wo ld, and he iden i ica ion o such egions is o subs an ial impo ance o wa e esou ces managemen . Howe e , coupled clima e models used o p edic u u e clima ic condi ions s ill exhibi a sp ead in he ep esen a ion o long- e m ends in soil mois u e and o he land wa e ela ed a iables (Guo & Di meye , 2006; Figu e 12.23 in Be g e al., 2017; Collins e al., 2013; Yuan & Qui ing, 2017). Compa ing he ou pu o nume ical models wi h obse a ions is c ucial o demons a e hei eliabili y and o es p edic i e capaci ies, bu measu emen s o wa e s o age changes a e di icul o ob ain. A classical RESEARCH ARTICLE 10.1029/2018JD029989 Key Poin s: • By compa ing e es ial wa e s o age ends om CMIP5 models and GRACE sa elli e da a, we iden i y ho spo egions o we ing and d ying • Model analysis e eals ha egional long- e m ends s a o eme ge om in e annual a ia ions a e 14 yea s • La ge model sp ead in wa e s o age ends demons a es impo ance o GRACE o cons ain esponse o he wa e cycle o clima e change Co espondence o: L. Jensen, lau a.jensen@hcu-hambu g.de Recei ed 14 NOV 2018 Accep ed 17 AUG 2019 Accep ed a icle online 29 AUG 2019 ©2019. The Au ho s. This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion License, which pe mi s use, dis ibu ion and ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed. JENSEN ET AL. 9808 Published online 3 SEP 2019 Ci a ion: Jensen, L., Eicke , A., Dobslaw, H., S acke, T., & Humph ey, V. (2019). Long‐ e m we ing and d ying ends in land wa e s o age de i ed om GRACE and CMIP5 models. Jou nal o Geophysical Resea ch: A mosphe es, 124, 9808–9823. h ps://doi.o g/ 10.1029/2018JD029989 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 app oach o he de e mina ion o TWS a basin scale is he in eg a ion o he wa e balance equa ion (p e- cipi a ion minus e apo anspi a ion minus uno ), see Rodell e al. (2004). Howe e , his is challenging on a global scale, since s eam low measu emen s a e spa se and e apo anspi a ion is gene ally di icul o measu e (Wa enbu ge e al., 2018). Especially, ends in wa e s o age canno be eco e ed well by his me hod due o biases in he wa e lux obse a ions (Hi schi & Sene i a ne, 2017). Complemen a y o con en ional me eo ologic obse a ions o a mosphe ic wa e luxes, he sa elli e mis- sion G a i y Reco e y And Clima e Expe imen (GRACE; Tapley e al., 2004) in ope a ion om 2002 o 2017 allowed o he i s ime he obse a ion o wa e s o age changes wi h global co e age om space. By e al- ua ing ela i e dis ance changes be ween wo spacec a a e y low al i udes o 400–500 km, ime a ia ions in he Ea h's g a i y ield a e mapped ha can be unambiguously ela ed o changes in TWS. Due o he indi ec obse a ion concep , GRACE essen ially senses wa e mass anomalies independen ly o hei su - ace exposu e and hus in eg a es all mass changes e ically om he su ace down o he deepes aqui e s. This unique capabili y o he g a ime ic me hod makes GRACE highly complemen a y o al e na i e adio- me ic sa elli e echniques o soil mois u e emo e sensing ha a e only sensi i e o changes in he op ew cen ime e s o soil (Do igo e al., 2015). GRACE mission da a ha e been used in a ious hyd ome eo olog- ical applica ions, o example, Famiglie i and Rodell (2013), and i is a ed among he op i e p io i ies o he u u e Ea h obse a ion capaci y by he mos ecen Na ional Ae onau ics and Space Adminis a ion decadal su ey (Commi ee on he Decadal Su ey o Ea h Science and Applica ions om Space e al., 2018). The successo mission GRACE-FO (Follow On), launched in May 2018, is expec ed o con inue his impo an obse a ional eco d o e he nex decades (Flech ne e al., 2016), which will acili a e he sep- a a ion be ween in e annual a iabili y and long- e m clima ological ends in TWS. Because he limi ed ime span o GRACE da a makes he iden i ica ion o clima e- ela ed signals s ill challenging, his s udy aims o in es iga e how GRACE TWS ends could (and should) be compa ed o model-de i ed ends. TWS as obse ed wi h GRACE has al eady been used o alida e bo h global hyd ological models (Döll e al., 2014; Eicke e al., 2014; Gün ne , 2008; Syed e al., 2008) and land su ace models (Scanlon e al., 2018; Zhang e al., 2017) which a e d i en by a p esc ibed me eo ological o cing. In his s udy, we ocus on in e - ac i ely coupled Ea h Sys em Models (ESMs) pa icipa ing in CMIP5 (Coupled Model In e compa ison P ojec Phase 5, Taylo e al., 2011). Compa ing GRACE ends wi h long- e m coupled clima e model p o- jec ions is challenging in mainly wo aspec s: (i) In con as o GRACE TWS (i.e., he ull in eg a ed wa e column, including all wa e ese oi s), TWS in he models is e lec ed ypically only by means o snow s o - age and soil mois u e. The ep esen a ion o he la e c i ically depends on he dep h o he soil column and he numbe o e ical laye s conside ed. In pa icula , cu en ESMs do no explici ly simula e g oundwa- e s o age changes. As g oundwa e -su ace in e ac ions play an impo an ole in he global hyd ological cycle, his poses an addi ional sou ce o unce ain y in long- e m model p ojec ions o we ing and d ying. (ii) Coupled uns in CMIP5 s a ing om p eindus ial condi ions and ex ending o e he whole his o ical pe iod un il he p esen day a e o ced wi h empo ally a iable sola adia ion, ae osols, CO2concen a- ions, and land use. Those expe imen s a e hus expec ed o ep oduce he clima e a iabili y in a s a is ical sense only. As a esul , di e en ealiza ions o he in e annual and decadal clima e a iabili y a e supe im- posed o e he clima ological ends so ha a di ec compa ison wi h he 14-yea GRACE TWS ime se ies only has limi ed explana o y powe . While a egional s udy o he Mississippi Basin (F eedman e al., 2014) showed easonably good ag eemen o he annual ampli ude o GRACE da a and a subse o CMIP5 models, Fasullo e al. (2016) ound he ends om his o ical CESM1-CAM5 uns compa ed o GRACE o be domi- na ed by in e nal a iabili y a he han by he o ced esponse. Di e en d i e s o TWS ends obse ed by GRACE we e in es iga ed by Rodell e al. (2018), who also made use o CMIP5 model p ecipi a ion p ojec- ions o a ibu e we ing and d ying endencies in some egions o clima e-d i en p ecipi a ion changes. To ou knowledge, an ex ensi e global compa ison o soil mois u e and snow ends modeled o e mo e han wo cen u ies (in he ollowing e e ed o as bicen ennial) agains GRACE obse a ions has ne e been conduc ed wi h an ensemble o models such as CMIP5. In esponse o hese challenges, we ocus in his s udy in pa icula on he co espondence o bicen ennial ends in TWS as simula ed by he majo i y o CMIP5 models and TWS endencies as obse ed by GRACE and in es iga e egions o ag eemen and disag eemen on we ing o d ying ends in models and sa elli e obse a ions. JENSEN ET AL. 9809 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Figu e 1. TWS ends om ITSG-G ace2018s (p elimina y) o he ime span Ap il 2002 o Augus 2016 (wi hou G eenland, S alba d, Gul Coas o Alaska, and An a c ica). S ippling indica es egions wi h nonsigni ican ends (𝛼=0.05). This pape is s uc u ed as ollows: Fi s , we compu e global maps o TWS ends om GRACE da a (sec ion 2) and CMIP5 models (sec ion 3) oge he wi h an e alua ion o he a iabili y among di e en models and wi hin his o ical and u u e ime spans. We es ima e he in luence o he di e en ime se ies leng hs o GRACE and models by means o wo model s udies using model TWS endencies om ime pe iods anging om 14 o mo e han 200 yea s (sec ion 4). The TWS end maps om GRACE and CMIP5 models a e subsequen ly compa ed (sec ion 5). Nex , we in es iga e ho spo and noncompliance egions o we ing and d ying ends ega ding hei unce ain y (sec ion 6), which migh be caused by model de ici s o na u al in e annual a iabili y and human impac s a ec ing he GRACE-de i ed ends. Sec ion 7 summa izes he esul s and add esses u u e wo k. 2. TWS T ends F om GRACE Da a To ob ain a global g id o obse ed TWS ends we use he ITSG-G ace2018s end Le el 2 da a (Maye -Gü e al., 2018), which was ob ained om es ima ing a long- e m mean g a i y ield model oge he wi h linea end and annual cycle om all a ailable GRACE Le el 1B RL03 da a. The ITSG-G ace2018s end used he e is a p elimina y e sion con aining Le el 1B da a o he ime span Ap il 2002 o Augus 2016 (∼14 yea s). I will be upda ed once he comple e ime se ies o Le el 1B RL03 da a (Ap il 2002 o June 2017) is a ailable. Howe e , o he end only mino changes a e expec ed by ex ending he ime se ies by less han 1 yea . The sphe ical ha monic coe icien s (Le el 2) o he end in g a i a ional po en ial a e gi en up o deg ee nmax =120 and a e pos p ocessed as ollows: The e ec o geocen e mo ion is aken in o accoun by aug- men ing he GRACE da a wi h he linea ends o deg ee 1 ha monic coe icien s p o ided by Swenson e al. (2008). The zonal Δc20 end coe icien is eplaced using a esul om Sa elli e Lase Ranging (Cheng e al., 2013). To educe mass ends o igina ing om glacial isos a ic adjus men (GIA), we sub ac a model om A e al. (2013) and o mi iga e he e ec o co ela ed noise a DDK4 il e (Kusche, 2007) is applied. We calcula e he TWS end on a 2◦×2◦geog aphical g id (Figu e 1) acco ding o ws(𝜆, 𝜃)= M 4𝜋R2𝜌w nmax ∑ n=1 n ∑ m=−n (2n+1) (1+k′ n)ΔcnmYnm(𝜆, 𝜃)(1) whe e 𝜆and 𝜃deno e he sphe ical coo dina es, Mand Ra e he mass and he adius o he Ea h, 𝜌w=1,000 kg/m3is he densi y o wa e , k′ ndeno e he Load Lo e Numbe s (Lambeck, 1988), Δcnm a e he il e ed sphe ical ha monic coe icien s o he g a i a ional po en ial, and Ynm(𝜆, 𝜃)a e he su ace JENSEN ET AL. 9810 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 sphe ical ha monic unc ions. Co esponding s anda d de ia ions o he TWS ends a e ob ained by a iance p opaga ion om ealis ic e o assump ions p o ided wi h he Δcnm coe icien s o he ITSG-G ace2018s end. The signi icance o he end can be es ed wi h a pa ame e es . The es ima ed end di ided by i s es ima ed s anda d de ia ion is compa ed o he c i ical alue o he no mal dis ibu ion o a ce ain signi icance le el 1−𝛼which we se o 95% in his s udy. Gene ally, he eliabili y o ends om GRACE is high. Among he solu ions o di e en GRACE p ocessing cen e s ends o e he same ime pe iod a e e y simila (Scanlon e al., 2018), e en i a di e en ep esen a ion (mascons ins ead o sphe ical ha monics) is chosen. Thus, selec ing ano he GRACE solu ion (e.g., om JPL o CSR) does no al e he indings o ou s udy (no shown). GRACE-de i ed ends migh no o igina e pu ely om TWS changes e e ywhe e, as esidual ec onic e ec s om GIA (Ca on e al., 2018), pos seismic de o ma ion a e la ge ea hquakes (Han e al., 2008, 2010), o esidual a mosphe ic mass a iabili y (Fagiolini e al., 2015) can o e lay TWS ends. Fu he mo e, leakage o signal in o neighbo ing g id cells due o il e ing and esidual noise ha could no be emo ed du ing il e ing migh also dis o TWS ends. As he GRACE TWS ends a e only calcula ed om 14 yea s o da a, he esul s can be domina ed by low- equency clima e a iabili y ela ed o El Niño–Sou he n Oscilla ion (Ni e al., 2018; Phillips e al., 2012), he sola cycle (Bha acha yya & Na asimha, 2005), he quasi-biennial oscilla ion and o he cou- pled clima e modes (G ay e al., 2018), and episodic e en s as olcanic e up ions (Iles e al., 2013), which may ei he conceal he long- e m end o p oduce a spu ious ansien end. App oaches o educe hese in e annual a iabili ies in he GRACE eco d a e cu en ly being discussed (e.g., Eicke e al., 2016). 3. TWS T ends F om CMIP5 Model Da a As CMIP5 models do no p o ide a s anda d ou pu a iable o o al wa e s o age, we use he sum o o al soil mois u e con en (m so) and su ace snow amoun (snw) as an app oxima ion o i . In he emaining pa o he pape we e e o his TWS app oxima ion as model TWS (mTWS). The mTWS di e s in se e al aspec s om GRACE-de i ed TWS: Soil mois u e laye s in ESMs ha e a dep h ha can a y widely be ween jus a ew and up o ens o me e s depending on he model and hus does no necessa ily cap u e he ull soil mois u e con en a e e y loca ion. Fu he mo e, g oundwa e and su ace wa e a e no explici ly included in mTWS as hese s a es a e gene ally no ep esen ed in CMIP5 models. Howe e , a ce ain ac ion o hese quan i ies migh be implici ly included in o al soil mois u e as he anspo o ocean and a mosphe e is limi ed and he wa e balance is la gely closed by mos o he models (Liepe & Lo, 2013). Mo eo e , his o ical CMIP5 uns do no con ain egional an h opogenic in e en ion o he han land use changes in hei se up (e.g., g oundwa e deple ion o dam building is no ep esen ed), whe eas GRACE obse a ions include hei consequences. The ep esen a ion o mTWS di e s om model o model due o di e en oo dep hs, numbe o soil laye s, and model physics (Huang e al., 2016). Snw also exhibi s la ge in e model di e ences in ep esen a ion (B u el-Vuilme e al., 2013). We he e o e no e ha mTWS o di e en models migh no be ully compa ible. A e adding mon hly m so and snw o each model, we conca ena e he co esponding mTWS simula- ions o he his o ical uns (1850–2005) and he RCP8.5 scena ios (2006–2100) o calcula e ends o ime spans ha go beyond he yea 2006. Fo hose models, whe e mo e han one un is a ailable, we calcula e he ensemble mean which we ega d as he mos obus ealiza ion o long- e m mTWS end es ima es. A e wa d, he mTWS alues a e emapped o a common 2◦×2◦geog aphical g id. A bicen ennial mTWS end map ( ime span Janua y 1861 o Decembe 2099, i.e., ea lies /la es common da e o all models o his o ical/RCP8.5 expe imen s) is calcula ed om he ime se ies o mTWS g ids o each model. Fo each g id cell he linea end is calcula ed by i ing a unc ion 𝑓( )=a+b· +c·cos(𝜔 )+d·sin(𝜔 )+e·cos(2𝜔 )+𝑓·sin(2𝜔 )(2) wi h pa ame e s o bias (a), linea end (b), annual and semiannual cycle (c,d,e, ) o he ime se ies by means o leas squa es adjus men . The s anda d de ia ion o he end is es ima ed om he pos i esidu- als. No e ha we exclude he glacia ed egions o G eenland, S alba d, Gul Coas o Alaska, and An a c ica, since no all models p ope ly ep esen glacie mass balance dynamics domina ing TWS in hose egions. JENSEN ET AL. 9811 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Figu e 2. Co ela ions o he bicen ennial mTWS end maps o 34 CMIP5 models. In o al 34 CMIP5 models p o ide a leas one un o m so and snw. Howe e , as some o hese 34 models a e ei he di e en e sions o he same model o a e uns wi h pa ly iden ical componen s (land su ace and/o a mosphe e model), i canno be assumed ha each model p oduces a comple ely independen es ima e o he m so and snw ields (Knu i e al., 2013). In o de o ob ain an unbiased mul imodel a e age mTWS end map and a eliable conclusion abou model consensus, we iden i y he independen models by compa ing he simila i y o mTWS end maps o all models. As a measu e o he simila i y o wo maps we use he Pea son p oduc -momen co ela ion coe icien 2calcula ed om he ec o ized maps, gi ing e e y land pixel o he 2◦×2◦g id equal weigh . As end ou lie s in single pixels can dis o he co ela ion coe icien we apply a simple h eshold o he mTWS end maps, excluding absolu e end alues abo e 23 mm/yea . This is he 2𝜎bounda y o he 14-yea GRACE TWS end (Figu e 1), hus i is e y unlikely ha bicen ennial ends abo e hese h eshold a e ealis ic. The co ela ions o he bicen ennial mTWS end maps (a e applying he 23 mm/yea h eshold) a e cal- cula ed o all 34 models and a anged in a ma ix (Figu e 2). De ailed in o ma ion and e e ences o he models lis ed in Figu e 2 a e gi en, o example, in Fla o e al. (2013) and a e no ei e a ed he e. As expec ed, models ha use common a mosphe e o land su ace componen s exhibi a e y high co ela ion. In o de o only conside models ha a e independen and o jus i y he applica ion o equal weigh o each model esul , in he emaining pa o he s udy we use only one ins ance om each g oup o models ha a e highly co ela ed ( 2>75%). In Figu e 2 he models ha a e excluded due o his h eshold a e deno ed in g ay on and he emaining 21 models a e highligh ed in bold on . The c i e ia o choosing a speci ic model among highly co ela ed models was based on i s es ima ed age (mos ecen publica ion), deg ee o special- iza ion (mos gene al), o spa ial esolu ion (closes o 2◦×2◦). Gene ally, a e excluding all bu one om he highly co ela ed models, he co ela ion among end maps om di e en models is e y low (mean 2=10%, maximum 2=67%) and o some pai s o models i is e en nega i e (minimum 2=−55%). This analysis demons a es he la ge inhomogenei y among CMIP5 models ega ding ends in mTWS. In o de o u he in es iga e model sp ead we de ine di e en ime spans (Table 1) o which we cal- cula e and discuss mTWS end maps in he ollowing. Fi s , he 21 models ha emain a e excluding highly co ela ed models, a e used o calcula e a median end map o he bicen ennial ime span Janua y JENSEN ET AL. 9812 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Table 1 No a ion o Di e en Time Spans Tha A e In es iga ed o mTWS T ends Time span No a ion Jan 1861 o Dec 2099 bicen ennial end Jan 1861 o Aug 2016 his o ical end Sep 2016 o Dec 2099 RCP8.5 end Jan 1986 o Dec 2035 50a endency Jan 1996 o Dec 2025 30a endency Ap 2002 o Aug 2016 14a endency 1861 o Decembe 2099 (Figu e 3a), ha is, o each geog aphical g id cell he median o he ends o all models is de e mined. We use he median ins ead o he a i hme ic unweigh ed mean because i is much less a ec ed by ou lie s and hus can be assumed o be a mo e obus es ima e o he end. How- e e , o iden i y nonsigni ican ends in he median map (s ippled egions in Figu e 3a) we ca y ou e o p opaga ion o he a i hme ic mean, because his is no s aigh o wa d o he median. Figu e 3a is no a ec ed by a model d i in mTWS, as ends om p eindus ial con ol simula ions (i.e., model uns only o ced wi h na u al, none ol ing a mosphe ic concen a ions) o he same CMIP5 models we e ound o be an o de o magni ude smalle and hus a e negligible (no shown). Acco ding o he 21 models, he la ges ends occu mainly in sou he n Eu ope and Tu key, in Cen al Ame ica and in he wes o No h Ame ica, in he no h o Sou h Ame ica and in he Himalaya egion. The clima ological ends de i ed he e a e in ag eemen wi h he esul s o a p e ious s udy (Be g e al., 2017) ha ocused on o al soil mois u e, e en hough signi ican di e ences a e p esen in high la i udes since mTWS also includes snowpack. To assess he eliabili y o he median mTWS ends, we compu e he le el o consensus o he 21 models, ha is, he numbe o models wi h he same bicen ennial end di ec ion o a gi en g id cell (Figu e 3b). The highe he consensus, he highe he ce ain y ha he ag eemen is no by chance, o example, i 15 o mo e o 21 models ag ee on he sign, he p obabili y ha his is jus chance is only 4% o less (Di meye e al., 2013). Hence, he highe he consensus in a g id cell, he mo e we can us he di ec ion o he end in his g id cell acco ding o he models. In many egions high consensus co esponds o la ge ends and ice e sa. Howe e , his is no alid e e ywhe e, meaning ha also he sign o small ends can be ep esen ed by a majo i y o models (e.g., India) and in e sely, he e migh be model disag eemen abou he di ec ion o la ge ends (e.g., No he n Russia). Fo he bicen ennial mTWS end, we ind 39% o he global land a ea o exhibi a d ying (30%) o we ing end (9%) ha is suppo ed by a leas 71% (15 o 21) o he models. These indings a e no ee o unce ain ies as he consensus map (Figu e 3b) migh be a ec ed by sys ema ic de ici s in CMIP5 models, such as in pa icula he lack o g oundwa e s o age in aqui e s a di e en dep h and hus e y di e en esidence imes (Pokh el e al., 2014). To in es iga e i mTWS ends as calcula ed o he bicen ennial ime span a e in p inciple al eady de ec able in obse a ional eco ds a ailable oday, we compu e (in addi ion o he bicen ennial ime span) mTWS ends o a his o ical ime span Janua y 1861 o Augus 2016 (un il he end o he GRACE ime span; Figu e 4a). Fo compa ison, also he mTWS ends o he RCP8.5 ime span Sep embe 2016 o Decembe 2099 a e displayed (Figu e 4b). O e all, we ind a simila pa e n o he his o ical and he RCP8.5 end (pa e n co ela ion o 55%), hough he his o ical end has a much smalle magni ude (only abou 20% o RCP8.5). Fu he mo e, he po ion o land a ea whe e he median ends a e no signi ican (s ippled a eas, 95% con idence le el) is la ge o he his o ical ime span han o he RCP8.5 ime span. Howe e , in 73% o he land a ea he his o ical end is al eady signi ican and in 68% i is in ag eemen wi h he RCP8.5 end map. In high-consensus egions (ag eemen o bicen ennial end sign in ≥71% o he models, Figu e 3b) o which we es ic he analysis in sec ion 5 and 6, he a ea o ag eemen be ween signi ican his o ical and RCP8.5 ends is 92%. This indica es ha in mos egions he cu en ends a e se o con inue in he same di ec ion and e en inc ease in he u u e, he eby sugges ing ha he p ocesses shaping he clima e change oo p in on TWS a e al eady ac ing oday. The consensus among he CMIP5 models is gene ally lowe o he his o ical ime span (Figu e 4c) han o he RCP8.5 ime span (Figu e 4d), which is ela ed o he ac ha s onge ends gene ally imply highe consensus and RCP8.5 is he scena io wi h he s onges clima e change signal. The pa e ns o he consensus maps a e simila o all h ee ime spans (bicen ennial, his- o ical, and RCP8.5), hus we in e ha egions o la ge model ag eemen a e la gely independen om he selec ed ime span ( o cen ennial ends). 4. In luence o Obse a ion Time Span F om Figu es 3 and 4 i can be concluded ha o cen ennial ime spans a empo ally s able pa e n o d ying and we ing ends exis s in he models. Howe e , we canno expec o eadily ind hese end pa e ns in JENSEN ET AL. 9813 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Figu e 3. (a) Median o bicen ennial mTWS end maps om 21 CMIP5 models (wi hou G eenland, S alba d, Gul Coas o Alaska, and An a c ica). S ippling indica es egions whe e he mean end is no signi ican ly di e en om ze o (𝛼=0.05). (b) Consensus map o bicen ennial mTWS ends om 21 CMIP5 models. Red colo s indica e ha ≥#models ag ee on a nega i e (i.e., d ying) end, blue colo s indica e ha ≥#models ag ee on a posi i e (i.e., we ing) end. a sho ime pe iod o only 14 yea s o which GRACE obse a ions a e a ailable. Fo sho ime pe iods, in e annual a ia ions may be domina ing he end es ima ion in many egions o he wo ld. To es ima e he in luence o he obse a ion ime span on he expec ed ag eemen wi h he bicen ennial end, we pe o m wo model s udies using endency maps o di e en ime spans calcula ed om he CMIP5 models. In he i s model s udy we in es iga e a e which ime span long- e m clima ic ends in mTWS migh be clea ly dis inguished om in e annual a ia ions. In he second model s udy we es ima e he deg ee o which e en a e long ime spans he obse ed ends migh s ill be in disag eemen wi h he long- e m clima ic end. In con as o he o he sec ions o he pape , whe e we ely on he ensemble means, o hese model s udies we only use one indi idual un ( 1i1p1) pe model in o de o p ese e in e annual a iabili y. This is impo an as na u al a ia ions would la gely a e age ou by calcula ing ensemble means. By using CMIP5 model ou pu o simula ing di e en ly long obse a ion ime spans, we p esume ha indi idual model uns ep esen na u al a iabili y ealis ically in e ms o ela i e magni ude, equency, and du a ion. JENSEN ET AL. 9814 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Figu e 4. ( op) Median o mTWS end maps om 21 CMIP5 models o (a) his o ical (Janua y 1861 o Augus 2016) and (b) RCP8.5 (Sep embe 2016 o Decembe 2099) ime span. S ippling indica es egions whe e he mean end is no signi ican ly di e en om ze o (𝛼=0.05). Please no e he di e en colo scales in (a) and (b). (bo om) Consensus maps o bicen ennial mTWS ends om 21 CMIP5 models o (c) his o ical and (d) RCP8.5 ime span. Fo he i s s udy we i ends in a leas squa es sense o ime spans o di e en leng hs anging om 14 o 100 yea s in s eps o 5 yea s wi h he cen e yea 2010. Examples o endency maps ob ained o he 50a, 30a, and 14a ime pe iods a e gi en in Figu e 5. The median endency maps o all 18 ime spans a e each co ela ed o he bicen ennial median end map (Figu e 5a). Fo he 14a obse a ion pe iod he global co ela ion is only 23%, bu wi h inc easing ime span i asymp o ically app oaches 100% (blue cu e in Figu e 5a). A e he 30a ime span he global co ela ion is 57% which is he same o de o simila i y ha we ind o he his o ical and RCP8.5 ime spans (55%). Thus we conclude ha a ound 30 yea s o TWS obse a ions would be he minimum ime o globally ob ain a TWS end compa able o long- e m model esul s. Howe e , e en hough he ag eemen be ween end pa e ns migh be low a 14a globally, his migh no be he case locally, o ins ance, when only conside ing egions ha exhibi s ong model ag eemen . When calcula ing he co ela ion o he 14a endency and he bicen ennial end only o g id cells wi h a model consensus o ≥71%, he co ela ion coe icien inc eases o 39% ( ed cu e in Figu e 5a), when addi ionally excluding nonsigni ican g id cells, i inc eases o 52% (yellow cu e in Figu e 5a). In his model s udy we e alua e he (global) spa ial pa e n co ela ion which only con ains limi ed in o - ma ion abou he ag eemen o ends o indi idual g id cells. This means ha hough his expe imen b ings ou wha o expec om he simila i y o he spa ial pa e ns, i does no p o ide he likelihood o a local mTWS endency compu ed om a ce ain ime span o ac ually ma ch he bicen ennial end in ha g id cell. As we a e in e es ed in egions whe e 14a GRACE TWS endencies ag ee wi h bicen ennial mTWS model ends and wan o a e he esul s wi h espec o wha o expec om his sho ime span, we pe - o m a second model s udy: Fo each o he 21 CMIP5 models we cu 22 slices o 14a mTWS da a wi h a dis ance o 5 yea s (cen e ed a ound he yea 1970) and es ima e 22 14a endencies. Fo each g id cell he ac ion o endencies ha ag ee o disag ee (in e ms o sign) wi h he bicen ennial mTWS end om ha pa icula model is calcula ed. Subsequen ly, he global mean o all ac ions and all models is compu ed. This p ocedu e is epea ed o di e en ime spans om 1 o 100 yea s in s eps o 5 yea s (Figu e 6). Acco d- ing o he models he p obabili y ha a 14a endency is in ag eemen wi h he long- e m end is on a e age 53%, which is sligh ly be e han andom chance. E en a e a cen u y he e is s ill a chance o 27% ha an indi idual endency does no ma ch he bicen ennial end e en hough he global pa e n co ela ion is al eady high wi h 87%. This indica es ha he e is na u al a iabili y in he models e en o e long ime pe i- ods o 100 yea s and mo e, which Laepple and Huybe s (2014) ound o be caused by sea su ace empe a u e JENSEN ET AL. 9815 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 Figu e 5. (a) Co ela ion o he mTWS endency maps o di e en ime spans (cen e yea 2010) wi h he bicen ennial mTWS end map. (b–d) Median o mTWS endency maps om 21 CMIP5 models o (b) 50a ime span, (c) 30a ime span, and (d) 14a ime span. Each ime span is cen e ed a ound he yea 2010. No e he di e en colo scales due o la ge a iabili y o sho e ime spans. S ippling indica es egions wi h nonsigni ican ends (𝛼=0.05). JENSEN ET AL. 9816 Jou nal o Geophysical Resea ch: A mosphe es 10.1029/2018JD029989 G ay, L. J., Ans ey, J. A., Kawa ani, Y., Lu, H., Osp ey, S., & Schenzinge , V. (2018). Su ace impac s o he Quasi Biennial Oscilla ion. A mosphe ic Chemis y and Physics,18(11), 8227–8247. h ps://doi.o g/10.5194/acp-18-8227-2018 Gün ne , A. (2008). Imp o emen o global hyd ological models using GRACE da a. Su eys in Geophysics,29(4), 375–397. h ps://doi.o g/ 10.1007/s10712-008-9038-y Guo, Z., & Di meye , P. A. (2006). E alua ion o he second global soil we ness p ojec soil mois u e simula ions: 1. In e model compa ison. Jou nal o Geophysical Resea ch,111, D22S02. h ps://doi.o g/10.1029/2006JD007233 Han, S.-C., Saube , J., & Lu hcke, S. (2010). Regional g a i y dec ease a e he 2010 Maule (Chile) ea hquake indica es la ge-scale mass edis ibu ion. Geophysical Resea ch Le e s,37, L23307. h ps://doi.o g/10.1029/2010GL045449 Han, S.-C., Saube , J., Lu hcke, S. B., Ji, C., & Polli z, F. F. (2008). Implica ions o pos seismic g a i y change ollowing he g ea 2004 Suma a-Andaman ea hquake om he egional ha monic analysis o GRACE in e sa elli e acking da a. Jou nal o Geophysical Resea ch,113, B11413. h ps://doi.o g/10.1029/2008JB005705 Hi schi, M., & Sene i a ne, S. I. (2017). Basin-scale wa e -balance da ase (BSWB): An upda e. Ea h Sys em Science Da a,9(1), 251–258. h ps://doi.o g/10.5194/essd-9-251-2017 Huang, Y., Ge be , S., Huang, T., & Lichs ein, J. W. (2016). E alua ing he d ough esponse o CMIP5 models using global g oss p ima y p oduc i i y, lea a ea, p ecipi a ion, and soil mois u e da a. Global Biogeochemical Cycles,30, 1827–1846. h ps://doi.o g/10. 1002/2016GB005480 Iles, C. E., Hege l, G. C., Schu e , A. P., & Zhang, X. (2013). The e ec o olcanic e up ions on global p ecipi a ion. Jou nal o Geophysical Resea ch: A mosphe es,118, 8770–8786. h ps://doi.o g/10.1002/jg d.50678 Knu i, R., Masson, D., & Ge elman, A. (2013). Clima e model genealogy: Gene a ion CMIP5 and how we go he e. Geophysical Resea ch Le e s,40, 1194–1199. h ps://doi.o g/10.1002/g l.50256 Kusche, J. (2007). App oxima e deco ela ion and non-iso opic smoo hing o ime- a iable GRACE- ype g a i y ield models. Jou nal o Geodesy,81(11), 733–749. h ps://doi.o g/10.1007/s00190-007-0143-3 Laepple, T., & Huybe s, P. (2014). Ocean su ace empe a u e a iabili y: La ge model-da a di e ences a decadal and longe pe iods. P oceedings o he Na ional Academy o Sciences,111(47), 16,682–16,687. h ps://doi.o g/10.1073/pnas.1412077111 Lambeck, K. (1988). Geophysical geodesy: The slow de o ma ions o he ea h. Ox o d [Ox o dshi e], New Yo k: Cla endon P ess: Ox o d Uni e si y P ess. h ps:// o e.nla.go .au/ e sion/21392411 Liepe , B. G., & Lo, F. (2013). CMIP5 upda e o ‘In e -model a iabili y and biases o he global wa e cycle in CMIP3 coupled clima e models’. En i onmen al Resea ch Le e s,8(2), 29401. h ps://doi.o g/10.1088/1748-9326/8/2/029401 Maye -Gü , T., Behzadpou , S., Ellme , M., K as, A., Klinge , B., & Zehen ne , N. (2018). ITSG-G ace2018—Mon hly, daily and s a ic g a i y ield solu ions om GRACE. GFZ Da a Se ices, h ps://doi.o g/10.5880/ICGEM.2018.003 Munday, C., & Washing on, R. (2018). Sys ema ic clima e model ain all biases o e Sou he n A ica: Links o mois u e ci cula ion and opog aphy. Jou nal o Clima e,31(18), 7533–7548. h ps://doi.o g/10.1175/JCLI-D-18-0008.1 Ni, S., Chen, J., Wilson, C. R., Li, J., Hu, X., & Fu, R. (2018). Global e es ial wa e s o age changes and connec ions o ENSO e en s. Su eys in Geophysics,39(1), 1–22. h ps://doi.o g/10.1007/s10712-017-9421-7 Phillips, T., Ne em, R. S., Fox-Kempe , B., Famiglie i, J. S., & Rajagopalan, B. (2012). The in luence o ENSO on global e es ial wa e s o age using GRACE. Geophysical Resea ch Le e s,39, L16705. h ps://doi.o g/10.1029/2012GL052495 Pokh el, Y. N., Fan, Y., & Miguez-Macho, G. (2014). Po en ial hyd ologic changes in he Amazon by he end o he 21s cen u y and he g oundwa e bu e . En i onmen al Resea ch Le e s,9(8), 84004. h ps://doi.o g/10.1088/1748-9326/9/8/084004 Pokh el, Y. N., Fan, Y., Miguez-Macho, G., Yeh, Pa J.-F., & Han, S.-C. (2013). The ole o g oundwa e in he Amazon wa e cycle: 3. In luence on e es ial wa e s o age compu a ions and compa ison wi h GRACE. Jou nal o Geophysical Resea ch: A mosphe es,118, 3233–3244. h ps://doi.o g/10.1002/jg d.50335 Rodell, M., Famiglie i, J. S., Chen, J., Sene i a ne, S. I., Vi e bo, P., Holl, S., & Wilson, C. R. (2004). Basin scale es ima es o e apo anspi- a ion using GRACE and o he obse a ions. Geophysical Resea ch Le e s,31, L20504. h ps://doi.o g/10.1029/2004GL020873 Rodell, M., Famiglie i, J. S., Wiese, D. N., Reage , J. T., Beaudoing, H. K., Lande e , F. W., & Lo, M.-H. (2018). Eme ging ends in global eshwa e a ailabili y. Na u e,557(7707), 651. h ps://doi.o g/10.1038/s41586-018-0123-1 Scanlon, B. R., Zhang, Z., Sa e, H., Sun, A. Y., Mülle Schmied, H., an Beek, L. P. H., e al. (2018). Global models unde es ima e la ge decadal declining and ising wa e s o age ends ela i e o GRACE sa elli e da a. P oceedings o he Na ional Academy o Sciences, 115(6), E1080–E1089. h ps://doi.o g/10.1073/pnas.1704665115 Swenson, S., Chambe s, D., & Wah , J. (2008). Es ima ing geocen e a ia ions om a combina ion o GRACE and ocean model ou pu . Jou nal o Geophysical Resea ch,113, B08410. h ps://doi.o g/10.1029/2007JB005338 Syed, T. H., Famiglie i, J. S., Rodell, M., Chen, J., & Wilson, C. R. (2008). Analysis o e es ial wa e s o age changes om GRACE and GLDAS. Wa e Resou ces Resea ch,44, W02433. h ps://doi.o g/10.1029/2006WR005779 Tapley, B. D., Be adpu , S., Wa kins, M., & Reigbe , C. (2004). The g a i y eco e y and clima e expe imen : Mission o e iew and ea ly esul s. Geophysical Resea ch Le e s,31, L09607. h ps://doi.o g/10.1029/2004GL019920 Taylo , K. E., S ou e , R. J., & Meehl, G. A. (2011). An o e iew o CMIP5 and he expe imen design. Bulle in o he Ame ican Me eo ological Socie y,93(4), 485–498. h ps://doi.o g/10.1175/BAMS-D-11-00094.1 T enbe h, K. E. (Ed.) (2010). Clima e sys em modeling. Camb idge: Camb idge Uni e si y P ess. OCLC: 845631695. Voss, K. A., Famiglie i, J. S., Lo, M., Linage, C. d, Rodell, M., & Swenson, S. C. (2013). G oundwa e deple ion in he Middle Eas om GRACE wi h implica ions o ansbounda y wa e managemen in he Tig is-Euph a es-Wes e n I an egion. Wa e Resou ces Resea ch, 49, 904–914. h ps://doi.o g/10.1002/w c .20078 Wa enbu ge , R., Sene i a ne, S. I., Hi schi, M., Chang, J., Ciais, P., De yng, D., e al. (2018). E apo anspi a ion simula ions in ISIMIP2a—E alua ion o spa io- empo al cha ac e is ics wi h a comp ehensi e ensemble o independen da ase s. En i onmen al Resea ch Le e s,13, 75001. h ps://doi.o g/10.1088/1748-9326/aac4bb Yuan, S., & Qui ing, S. M. (2017). E alua ion o soil mois u e in CMIP5 simula ions o e he con iguous Uni ed S a es using in si u and sa elli e obse a ions. Hyd ology and Ea h Sys em Sciences,21(4), 2203–2218. h ps://doi.o g/10.5194/hess-21-2203-2017 Zhang, L., Dobslaw, H., S acke, T., Gün ne , A., Dill, R., & Thomas, M. (2017). Valida ion o e es ial wa e s o age a ia ions as simula ed by di e en global nume ical models wi h GRACE sa elli e obse a ions. Hyd ology and Ea h Sys em Sciences,21(2), 821–837. h ps:// doi.o g/10.5194/hess-21-821-2017 JENSEN ET AL. 9823