Impact of the 2015/2016 El Niño on the terrestrial carbon cycle constrained by bottom-up and top-down approaches

Evaluating the response of the land carbon sink to the anomalies in temperature and drought imposed by El Niño events provides insights into the present-day carbon cycle and its climate-driven variability. It is also a necessary step to build confidence in terrestrial ecosystems models' response to the warming and drying stresses expected in the future over many continents, and particularly in the tropics. Here we present an in-depth analysis of the response of the terrestrial carbon cycle to the 2015/2016 El Niño that imposed extreme warming and dry conditions in the tropics and other sensitive regions. First, we provide a synthesis of the spatio-temporal evolution of anomalies in net land–atmosphere CO2 fluxes estimated by two in situ measurements based on atmospheric inversions and 16 land-surface models (LSMs) from TRENDYv6. Simulated changes in ecosystem productivity, decomposition rates and fire emissions are also investigated. Inversions and LSMs generally agree on the decrease and subsequent recovery of the land sink in response to the onset, peak and demise of El Niño conditions and point to the decreased strength of the land carbon sink: by 0.4–0.7 PgC yr−1 (inversions) and by 1.0 PgC yr−1 (LSMs) during 2015/2016. LSM simulations indicate that a decrease in productivity, rather than increase in respiration, dominated the net biome productivity anomalies in response to ENSO throughout the tropics, mainly associated with prolonged drought conditions. This article is part of a discussion meeting issue ‘The impact of the 2015/2016 El Niño on the terrestrial tropical carbon cycle: patterns, mechanisms and implications’.

present-day carbon cycle and its climate-driven variability. It is also a necessary step to build confidence in terrestrial ecosystems models' response to the warming and drying stresses expected in the future over many continents, and particularly in the tropics. Here we present an in-depth analysis of the response of the terrestrial carbon cycle to the 2015/2016 El Niñ o that imposed extreme warming and dry conditions in the tropics and other sensitive regions. First, we provide a synthesis of the spatiotemporal evolution of anomalies in net land-atmosphere CO 2 fluxes estimated by two in situ measurements based on atmospheric inversions and 16 land-surface models (LSMs) from TRENDYv6. Simulated changes in ecosystem productivity, decomposition rates and fire emissions are also investigated. Inversions and LSMs generally agree on the decrease and subsequent recovery of the land sink in response to the onset, peak and demise of El Niñ o conditions and point to the decreased strength of the land carbon sink: by 0.4-0.7 PgC yr 21 (inversions) and by 1.0 PgC yr 21 (LSMs) during 2015/2016. LSM simulations indicate that a decrease in productivity, rather than increase in respiration, dominated the net biome productivity anomalies in response to ENSO throughout the tropics, mainly associated with prolonged drought conditions. This article is part of a discussion meeting issue 'The impact of the 2015/2016 El Niñ o on the terrestrial tropical carbon cycle: patterns, mechanisms and implications'.

Introduction
The global terrestrial CO 2 sink has increased steadily in the past decades but presents high year-to-year variations that, in turn, dominate inter-annual variability (IAV) in the atmospheric CO 2 growth rate [1]. As the atmospheric CO 2 growth rate is highly correlated with tropical temperature [2], IAV in the land sink has been mainly attributed to tropical forests [2], but semi-arid ecosystems appear to be increasingly important [3][4][5].
The El Niñ o/Southern Oscillation (ENSO) is an atmosphere-ocean variability pattern that drives temperature and rainfall variations in the tropics, with teleconnections that extend worldwide [6]. El Niñ o events strongly reduce the global land sink by up to 2PgC [7], leading to high atmospheric CO 2 growth rates [1]. El Niñ o events promote drought conditions in the Amazon forest, leading to increased tree mortality and reduced carbon storage [8,9] and widespread fires, particularly in southeast Asia [10]. ENSO impacts extend beyond the tropics, controlling IAV in sub-tropical ecosystem productivity [11], especially water-limited ecosystems in the Southern Hemisphere [3,4,12]. Most Coupled Model Intercomparison Project Phase 5 (CMIP5) models projected a two-fold increase in the frequency of extreme El Niñ o events in the future decades [13], associated with intensification of ENSOrelated anomalies in the carbon cycle [14]. However, nonlinear ENSO dynamics found in observations and one model might imply suppressed extreme El Niñ o events under warming [15].
Additionally, ENSO affects key regions and processes that are sources of uncertainty in future carbon cycle projections [3,16]. It is still unclear if temperature [2] or water-availability [3,9,11] drive ecosystems' response to ENSO, and how gross primary productivity (GPP) and terrestrial ecosystem respiration (TER) contribute to IAV. Analysis of model ensembles suggests that because water availability enhances both GPP and TER, its effects are cancelled out, and only the temperature signal emerges [2,5]. Jung et al. [5] also showed that water availability is the primary driver of carbon fluxes at the local scale, but anomalies tend to compensate spatially, so temperature emerges as a stronger driver with increasing spatial aggregation.
More generally, IAV in the carbon cycle is still not well understood, and neither data-driven models [17] nor Earth-System Models [18] capture its amplitude. In the 2017 Global Carbon Budget [1], land-atmosphere CO 2 fluxes from land-surface models (LSMs, bottom-up) forced with observed climate and land-use change (LUC) show good agreement with estimates from atmospheric transport model inversions (top-down) for global totals but differ at regional or zonal scale [1]. The 2015/ 2016 El Niñ o is especially interesting, as 2015 registered record atmospheric CO 2 growth rate in spite of widespread recordbreaking greening and stabilization of fossil-fuel emissions [1,19]. The 2015/2016 El Niñ o therefore provides a good study case to understand the response of ecosystems to warm and dry extremes potentially concurrent with global vegetation greening.
The strong El Niñ o event started around May 2015 and persisted until mid-2016, being the strongest event since the 1950s [20]. Record-breaking temperatures and drought were registered in the Amazon from October 2015 onwards. The drought extent in the Amazon was comparable to 1997/ 1998 but the extreme temperatures led to an exacerbation of dryness, with extreme drought conditions affecting double the extent of 1997/1998 [20].
According to LeQuéré et al. [1], the atmospheric CO 2 growth rate in 2015 and 2016 was 1.6 and 1.5 PgC yr 21 higher than during the 2011-2016 period, respectively, yet CO 2 emissions from fossil fuel and LUC combined were only 0.2-0.4 PgC yr 21 above the previous 5-year mean. Ocean uptake was estimated to be slightly larger (0.2 PgC yr 21 ) in 2015/2016 than the 2010-2014 average. Table 1 shows the residual sink needed to close the global carbon budget: the terrestrial CO 2 uptake had to be reduced by 1.4 PgC yr 21 in 2015 and by 1.5 PgC yr 21 in 2016. In the same period, but using the year of 2011 as a reference, Liu et al. [21] reported much higher losses of CO 2 over the pantropical regions in 2015 alone (2.5 PgC). Contrary to the 1997/1998 event, the anomaly in the land sink during 2015/2016 does not appear to be associated with major fire emissions. Although the development of El Niñ o coincided with enhanced fire activity in Southeast Asia, fire emissions in the region were reported to be only half of the emissions during the previous El Niñ o in 1997/1998, following rainfall return in November 2015 [22]. GFED4.1s [23] reports fire emissions 0.3 PgC yr 21 higher than the previous 5 years in 2015, but lower by 0.1 PgC yr 21 in 2016 (table 1).
Here we quantify the response of the terrestrial carbon cycle to El Niñ o in 2015/2016 using multiple data-based and modelled datasets. We track the evolution of anomalies in the net land -atmosphere CO 2 flux during the development and decline of the 2015/2016 El Niñ o estimated by two atmospheric transport model CO 2 inversions [24,25] and compare them with the net terrestrial CO 2 uptake and its component fluxes (gross primary productivity (GPP), total ecosystem respiration (TER), fire) simulated by 16 LSMs in the latest TRENDY intercomparison project (v6,

Material and methods (a) Atmospheric CO 2 inversion fluxes
Here we use three observation-based datasets of net landatmosphere surface fluxes: the Copernicus Atmosphere Monitoring Service (CAMS) atmospheric inversion (henceforth simply 'inversion') version 16r1 [24,43], and the Jena CarboScope inversion (update of [25,44] compare with Rödenbeck et al. [45]) versions s76_v4.1 and s04_v4.1 (CarboScope76 and CarboScope04 henceforth). The inversions provide terrestrial (and oceanic) surface CO 2 fluxes, CAMS weekly fluxes at 1.98latitude Â 3.758longitude resolution, and CarboScope daily fluxes at 48latitude Â 58longitude resolution. CAMS 16r1 uses 119 atmospheric stations over the different time frames for which they provide data, starting in 1979. CarboScope76 (CarboScope04) uses 10 (59) stations continuously available throughout 1976-2016 (2004-2016). All inversions are regularized by a priori information. CAMS uses climatological natural fluxes and time-varying ocean, wildfire and fossil-fuel fluxes with error correlation lengths of 4 weeks and 500 km (1000 km) over land (ocean) [46]. CarboScope uses a zero land prior, and a priori correlations of about 1600 km in longitude direction, 800 km in latitude direction and about 3 weeks. The inversions further differ in the transport model used, and other characteristics. Thus, they provide a range of uncertainty for observation-based top-down CO 2 flux estimates [19]. We focus on the 38-year period from 1979 until 2016 and calculate monthly anomalies of net land-atmosphere fluxes by subtracting the mean seasonal cycle and the monthly long-term trend (using a simple linear fit). We aggregate the inversion results over large regions (global terrestrial surface and tropical band between 238S and 238N), as flux estimates from inversions carry smaller relative uncertainties on the larger spatial scale [47].

(b) Land-surface models
LSMs simulate the key energy, hydrological and carbon cycle processes in ecosystems, allowing insights on the mechanisms controlling anomalies in land -atmosphere CO 2 fluxes and their drivers. The TRENDY intercomparison project coordinated historical LSM simulations and compiled outputs of CO 2 fluxes among other variables [42]. We use 16 LSMs from the latest TRENDYv6 simulations [1] (table 2), which provide monthly CO 2 fluxes during 1860 -2016. In TRENDYv6 S3 simulations, models are forced by historical data of (i) atmospheric CO 2 concentrations, (ii) climate observations from CRU-NCEP v8 [48,49] and (iii) human-induced land-cover changes and management from the HYDE [50,51] and the Land-Use Harmonization LUH2 v2 h [52] datasets (extended to 2016 as described in [1]). We analyse monthly values of net biome productivity (NBP), GPP, total ecosystem respiration (TER) and fire emissions simulated by the models (only 7 models) and annual leaf-area index (LAI, 12 models). NBP corresponds to the simulated net atmosphere-land flux ( positive sign for a CO 2 sink) and is comparable to top-down estimates of net land-atmosphere CO 2 fluxes, although the latter include lateral C fluxes (the land -ocean transport of C in freshwater and coastal areas and C fluxes due to trade/import export) [1,53] not simulated by the models. However, we focus on flux anomalies that should not be substantially affected by lateral fluxes because they are assumed to vary little between years. To produce a spatially consistent ensemble, model outputs were remapped to a common regular 18 Â 18 grid. The model data were selected for the 38-year long period 1979-2016, common to inversions.

(c) Satellite-based data
We compare anomalies from inversions and LSMs with two remote-sensing datasets that provide proxies for ecosystem activity and a satellite-based GPP product. Table 1. Global carbon budget during 2015, 2016 from the latest Global Carbon Project global carbon budget estimates (GCB2017v1.2, [1]). Annual atmospheric CO 2 growth rate (G ATM ), fossil fuel and LUC emissions (E FF and E LUC , respectively) and the total sinks partitioned into ocean and land fluxes. The numbers in brackets indicate the corresponding anomaly relative to the previous 5-year period. The land sink is estimated here as the residual from the global carbon budget (i.e. E FF þ E LUC 2 G ATM 2 O). Fire emission anomalies from GFED4.1s (1997 -2016) VISIT Y Y [41] rstb.royalsocietypublishing.org Phil. Trans. R. Soc. B 373: 20170304 LAI is defined as the one-sided green leaf area per unit ground area in broadleaf canopies and as one-half of the green needle surface area in needleleaf canopies, which depicts the greenness of vegetation. We used Collection 6 Terra and Aqua MODIS LAI products (MOD15A2H and MYD15A2H) [54,55]. The original datasets were available as 8-day composites in 500 m sinusoidal projection. We checked the quality flags (clouds, aerosols, etc.) to get high-quality LAI as described by Samanta et al. [56]. The original data were re-projected onto a 1/128 Â 1/128 grid by averaging the high-quality LAI. After that, the two LAI datasets were combined to bi-monthly timesteps by taking the mean of LAI values in each 8-day composite, weighted by the number of days that each 8-day composite locates in the specific half-month window. Finally, the annual average LAI and its anomaly relative to the record period (2000-2016) were calculated for each pixel. Anomalies in LAI reflect changes in the canopy leaf density and can therefore track plant stress response to drought.
Cheng et al. [57] used ground-based and remotely sensed land and atmospheric observations, combined with water use efficiency (WUE) model and evapotranspiration data from global land evaporation Amsterdam model (GLEAM), to calculate global annual GPP between 2000 and 2016 at 0.5 Â 0.58 resolution. The WUE model was developed by upscaling leaf WUE directly and considers the controls of vapour pressure deficit and physiological functioning on WUE. The model has been derived independently from GPP and evapotranspiration data, and therefore, can be used to evaluate simulated GPP.
Vegetation optical depth (VOD) is an estimate of the vegetation extinction effects on microwave radiation and increases with increasing vegetation density, being therefore a good proxy of biomass [58]. Brandt et al. [59] have shown that the new L-band soil moisture and ocean salinity (SMOS) VOD (L-VOD) retrieved from the SMOS-IC algorithm (Version V105 [60]) relates almost linearly to biomass and is thus relevant to monitor carbon stocks at continental scales. In this algorithm, no auxiliary data (either from atmospheric models or remote sensing optical observations) are used, except for surface temperature data from European Centre for Medium-Range Weather Forecasts (see [58,60] for more details). As L-VOD shows a strong relationship with aboveground biomass stocks, the time-derivative of L-VOD can be directly related to variations in biomass, and thus comparable with the aboveground component of NBP.

Results
(a) Global and tropical net biome productivity anomalies

(d) Driving processes
For further insight into the processes driving the land sink response to El Niñ o, we analyse the seasonal evolution of GPP anom and TER anom simulated by the LSMs (figure 3d-f) during 2015/2016. Electronic supplementary material, figure S4 additionally shows spatial GPP anom estimated by the MMEM from Q1 to Q8. LSMs indicate an increase in GPP during the first half of 2015 mainly in the extra tropics (consistent with the record greening that year [19,62]). Only a few regions in southern Africa and the Sahel and in Australia registered negative GPP anom already in Q1 and Q2 (electronic supplementary material, figure S3). The MMEM shows negative global GPP anom during the abrupt onset of El Niñ o (Q3), but also large spread, while negative GPP anom and spread in the tropics are still relatively small for Q3. Most LSMs estimate a strong negative global and tropical GPP anom during the peak of El Niñ o (Q4 and Q5), mostly over the Amazon and eastern Brazil, as well as extra-tropical southern Africa and Australian regions (electronic supplementary material, figure S4). LSMs simulate weak negative GPP anom     figure S5), except for central Africa (where above-average GPP is simulated). The spatiotemporal evolution of simulated TER anom appears, thus, to be mainly dominated by changes in GPP. The spatio-temporal evolution of simulated GPP anom mentioned above followed the progressive drying as El Niñ o developed (evaluated using a multi-scalar drought index at 6-month time-scale; electronic supplementary material, figure S6). The peak of El Niñ o in Q4 and Q5 corresponded to increasing intensity and spatial extent of drought conditions, affecting almost all tropical regions in South America, Asia and Australia and persisting until Q6 or even Q7 (South America and Australia). Even though in South America the peak of drought coincided with widespread negative GPP anom , the largest decreases in productivity are observed in typically dry regions, while humid areas (central Amazon) show smaller anomalies in productivity and recover faster (with positive anomalies in Q7). In Africa, the dipole of wet conditions in central tropics versus strong dryness in the south largely matches that of GPP anom .

(e) Comparison with satellite-based data
We evaluate whether simulated anomalies in vegetation status and productivity are consistent with LAI from MODIS and GPP derived from satellite data using a wateruse efficiency model (GPP-WUE), shown in figure 4. We further evaluate changes in vegetation-optical depth as a proxy for changes in aboveground biomass. LSMs estimate widespread negative LAI anomalies in most of the tropics in both years, consistent with MODIS LAI. LSMs simulate positive LAI anom for the humid forests in Africa, where MODIS LAI anom shows more heterogeneity. Both MODIS and simulated LAI report an amplification of negative anomalies in 2016, also extending to parts of the Amazon.
The regions with strongest LAI decrease roughly coincide with those regions where below-average anomalies are found in both WUE-derived and simulated GPP: dry forests in tropical South America, the southern section of Africa and the Sahel, continental Southeast Asia and northern Australia. The agreement between WUE-GPP anom and MMEM GPP anom is better in 2015 than in 2016, though. In humid forests in Africa, WUE-GPP shows generalized negative anomalies in 2016, while LSMs simulate positive GPP anom .
The L-VOD index used here is more sensitive to the whole vegetation layer than other indices, which are more sensitive to the upper part of the canopy [59]. Even though L-VOD decrease (biomass reduction) is registered in the dry forests and savannahs of South America as in LAI and GPP, positive

Discussion
Our results show that the LSMs in TRENDYv6 can reproduce IAV patterns of the global terrestrial C-sink very close to the anomaly in the residual sink from GCB2017 and within the spread of atmospheric transport model inversions. That study pointed to distinct continental-scale processes explaining anomalies in CO 2 fluxes: GPP decrease in tropical America, TER increase in Africa and fire activity in Asia. Our results agree on the dominant role of GPP decrease in South America during El Niñ o. However, we find strong intra-continental heterogeneity, with strongest negative GPP anom in dry forests and savannahs, consistent with previous studies showing the dominant role of semi-arid ecosystems in controlling carbon cycle sensitivity to ENSO [3,4]. Neither study does a perfect attribution of TER: TER in [21] is calculated as a residual term and might therefore be affected by errors in their NBP, GPP and fire emission estimates; at the same time, LSMs do not represent realistically the sensitivity of TER to precipitation [64]. Contrary to [21], LSMs indicate that tropical TER also decreased overall, probably because of the reduced substrate of TER or inhibition of decomposition due to drought. In Africa, the LSMs simulate a dipolar pattern during the peak of El Niñ o for both GPP anom and TER, with an increase in the 08 -208S region but a decrease in both variables further south. WUE-GPP shows similar results for 2015, but points to generalized negative GPP anom in 2016. The decrease in TER in regions with decreased GPP may indicate a strong coupling of TER with biomass production in LSMs, as spatio-temporal anomalies in GPP and TER are mainly in phase, as noted previously [5].
The subset of LSMs that simulate fires shows a moderate increase in emissions (global average of 0.2 GtC for the 2 years), but significantly lower than the Liu et al. [21] estimate of fire emission increase of 0.4 GtC for South Asia only. This difference may be due to the lack of peat fires in LSMs but is hard to reconcile with the lower GFED4.1s estimate of global fire emissions (electronic supplementary material, figure S3). LSMs could show too little sensitivity of TER and fires to climate variability, several models sharing similar parametrizations to represent soil decomposition response to temperature and water stress for example. Conversely, the Liu et al. [21] study uses sun-induced chlorophyll fluorescence as an indirect measure of GPP and carbon monoxide (CO) concentrations as a proxy for fires. How these relationships or systematic errors in assimilated total column CO 2 retrievals vary between normal and El Niñ o years is still unclear.
Even though below-average GPP was registered in the Amazon (especially in 2016) in both LSM simulations and WUE-GPP, the strongest decreases in GPP occur in the tropical dry forest and savannahs in South America, southern Africa and northern Australia. This points to a predominant role of water availability in the observed response to the 2015/2016 El Niñ o and is consistent with previous studies [3,4,9,39,59]. Indeed, the spatio-temporal evolution of simulated tropical GPP decrease during the onset and peak of the 2015/2016 El Niñ o follows the progressive increase in dryness (electronic supplementary material, figure S6). Additionally, LSMs indicate that dry forests and semi-arid biomes respond more strongly than humid ones to similar drought conditions and also point to a faster recovery of the humid Amazon forest in the second half of 2016 (electronic supplementary material, figure S4), when drought conditions started to become more moderate (electronic supplementary material, figure S6).
It is worth pointing out that the good agreement between LSMs and inversions or satellite-based observations is especially true for the MMEM, while individual models may show substantially different regional response over the course of the 2015/2016 event, although most individual LSMs show anomalies consistent with the MMEM across the tropics (electronic supplementary material, figure S7). Since all models use the same climate and land-use forcing, the differences in model responses arise because of the different parametrizations of carbon cycle processes, resulting in different model sensitivities to the increasingly warm and dry conditions observed until the peak of El Niñ o. The added value of using MMEM is recognized Earth system modelling, and several examples exist of applications in which combined information from several models is superior to results from any single model [65]. In the climate community, the diversity amongst models is considered a healthy aspect and provides a basis for estimating uncertainty [66].

Conclusion
We show that the LSM ensemble reproduces the spatial and temporal impacts of the 2015/2016 El Niñ o on the terrestrial C-sink within the inversions' range. We find that the decrease in the global terrestrial sink during El Niñ o in 2015/2016 can be mainly explained by decreased tropical GPP, in response to the ENSO-related drought in transitional to semi-arid regions, with a secondary role of the increase in fires and ecosystem respiration. It is still unclear whether TER plays an important role in controlling NBP anom during El Niñ o events. Our results agree with recent work highlighting the control of NBP by water availability [3,5]. However, this agreement might be ENSO event-dependent, as we found larger disagreement between inversions and LSMs in 1997/ 1998 than in 2015/2016. Understanding how terrestrial biogeochemical processes contribute to the emergent response of ecosystems to warming and drying during El Niñ o events is crucial to comprehend the vulnerability of land ecosystems to future changes in climate in the tropics and other sensitive regions.
Data accessibility. CO 2 fluxes from the CAMS atmospheric inversion are freely available at http://atmosphere.copernicus.eu/. CarboScope datasets are available at http://www.bgc-jena.mpg.de/Carbo-Scope/. The monthly time series of global and tropical CO 2 fluxes are provided for each inversion in electronic supplementary material. Outputs from land surface models from the TRENDYv6 project used in this study are provided in electronic supplementary material. These are the time series of global/tropical average monthly NBP, GPP, TER and fire emissions from each individual model, as well as the annual and seasonal gridded anomalies of the MMEM NBP, GPP, TER and LAI. Annual and seasonal anomaly NBP maps from inversions, and annual NBP maps from individual models are also provided. The full TRENDYv6 dataset including other outputs is available, subject to the individual modelling groups' agreement, via a request to S. Sitch (s.a.sitch@exeter.ac.uk). The results from the Global Carbon Budget 2017 are available for download at http:// www.globalcarbonproject.org/carbonbudget/17/data.htm. Annual anomalies of WUE-GPP are provided in electronic supplementary material. The GFED4.1s fire emission database is publicly available at http://www.globalfiredata.org/data.html. MODIS data are freely available at https://lpdaac.usgs.gov/dataset_discovery/ modis/modis_products_table. The L-VOD data can be accessed at the Centre Aval de Traitement des Données SMOS (CATDS): ftp://ext-catds-cecsm:catds2010@ftp.ifremer.fr/Land_products/L3_ SMOS_IC_Soil_Moisture/Seasonal_Averages/. The Standardized Precipitation-Evapotranspiration Index is freely available at http:// spei.csic.es/.