This study makes use of the CO2-responsive version of the ISBA LSM included in the open-access SURFEX modelling platform of Météo-France (Masson et al., 2013). The most recent version of SURFEX (version 8.1) is used with the “NIT” biomass option for ISBA. The latter simulates the diurnal cycle of water and carbon fluxes, plant growth, and key vegetation variables like LAI and above-ground biomass on a daily basis. It can be coupled to the CTRIP river-routing model in order to simulate streamflow. In this version of ISBA, a single-source energy budget of a soil–vegetation composite is computed. Also, the ISBA parameters are defined for 12 generic land surface patches, which include nine plant functional types (needle leaf trees, evergreen broadleaf trees, deciduous broadleaf trees, C3 crops, C4 crops, C4 irrigated crops, herbaceous, tropical herbaceous and wetlands), bare soil, rocks, and permanent snow and ice surfaces. A more comprehensive model description can be found in Masson et al. (2013).

ISBA accounts for the atmospheric CO2 concentration on stomatal aperture (Calvet et al., 1998, 2004; Gibelin et al., 2006). Also, photosynthesis and its coupling with stomatal conductance on a leaf level are accounted for. The vegetation net assimilation of CO2 is estimated and used as an input to a simple vegetation growth submodel able to predict LAI: photosynthesis drives the dynamic evolution of the vegetation biomass and LAI variables in response to atmospheric and climate conditions. During the growing phase, enhanced photosynthesis corresponds to a CO2 uptake, which leads to vegetation growth. In contrast, lack of photosynthesis leads to higher mortality rates. The GPP is defined as the carbon uptake while the ecosystem respiration (RECO) is the release of CO2, the difference between these two quantities being the net ecosystem CO2 exchange (NEE). Evaporation due to (i) plant transpiration, (ii) liquid water intercepted by leaves, (iii) liquid water contained in top soil layers and (iv) the sublimation of snow and soil ice are combined to represent the total evaporative flux.

The ISBA 12-layer explicit snow scheme (Boon and Etchevers, 2001; Decharme et al., 2016) and its multilayer soil diffusion scheme (ISBA-Dif) are used. The later is based on the mixed form of the Richards equation (Richards, 1931) and explicitly solves the one-dimensional Fourier law. It also incorporates soil freezing processes developed by Boone et al. (2000) and Decharme et al. (2013). The total soil profile is vertically discretized; both the temperature and moisture of each soil layer are computed according to their textural and hydrological characteristics. The Brookes and Corey model (Brooks and Corey, 1966) determines the closed-form equations between the soil moisture and the soil hydrodynamic parameters, including the hydraulic conductivity and the soil matrix potential (Decharme et al., 2013). The default discretization with 14 layers over 12 m depth is used. The lower boundary of each layer being: 0.01, 0.04, 0.1, 0.2, 0.4, 0.6, 0.8, 1, 1.5, 2, 3, 5, 8 and 12 m deep (see Fig. 1 of Decharme et al., 2011). Amounts of clay, sand and organic carbon in the soil determine the thermal and hydrodynamic soil properties (Decharme et al., 2016). They are taken from the Harmonized World Soil Database (HWSD; Wieder et al., 2014). As for hydrology, the infiltration, surface evaporation and total runoff are accounted for in the soil water balance. The infiltration rate defines the discrepancy between the surface runoff and the throughfall rate. The later being defined as the sum of rainfall not intercepted by the canopy, dripping from the canopy (i.e. interception reservoir) and snow melt water. The soil evaporation affects only the superficial layer (top 1 cm) and is proportional to its relative humidity. Transpiration water from the root zone (the region where the roots are asymptotically distributed) follows the equations in Jackson et al. (1996). Canal et al. (2014) provide more information on the root density profile.

Both the surface runoff (the lateral subsurface flow in the topsoil) and a free drainage condition at the bottom soil layer contribute to ISBA total runoff. The Dunne runoff (i.e. when no further soil moisture storage is available) and lateral subsurface flow from a subgrid distribution of the topography are computed using a basic TOPMODEL approach. The Horton runoff (i.e. when rainfall has exceeded infiltration capacity) is estimated from the maximum soil infiltration capacity and a subgrid exponential distribution of the rainfall intensity.

CTRIP is driven by three prognostic equations corresponding to (i) the groundwater, (ii) the surface stream water and (iii) the seasonal floodplains. Streamflow velocity is computed using the Manning formula as described in Decharme et al. (2010). When the river water level overtops the river-bank, it fills up the floodplain reservoir which empties when the water level drops below this threshold (Decharme et al., 2012). Occurrence of flooding impacts the ISBA soil hydrology through infiltration, and it also influences the overlying atmosphere via free surface-water evaporation and precipitation interception. The groundwater scheme is based on the two-dimensional groundwater flow equation for the piezometric head (Vergnes and Decharme, 2012). Its coupling with ISBA enables accounting for the presence of a water table under the soil moisture column. It allows for upward capillary fluxes into the soil (Vergnes et al., 2014). CTRIP is coupled to ISBA through OASIS-MCT (Voldoire et al., 2017). Once a day, ISBA provides CTRIP with updates on runoff, drainage, groundwater and floodplain recharges, and CTRIP feedbacks to ISBA the water table fall or rise, floodplain fraction, and flood potential infiltration. The current CTRIP version consists of a global streamflow network at 0.5∘ × 0.5∘ spatial resolution.

USCRN is a network of climate-monitoring stations maintained and operated by the National Oceanic and Atmospheric Administration (NOAA). It aims at providing climate-science-quality measurements of air temperature and precipitation. To increase the network's capability of monitoring soil processes and drought, soil observations were added to USCRN instrumentation. At each USCRN station in the conterminous United States in 2011, the USCRN team completed the installation of triplicate-configuration soil moisture and soil temperature probes at five standard depths (5, 10, 20, 50 and 100 cm) as prescribed by the World Meteorological Organization. The 111 stations present data between 2009 and 2016. Stations provide data at an hourly time step. Similar to a prior study, data sets potentially affected by frozen conditions were masked out using an observed temperature threshold of 4 ∘C (e.g. Albergel et al., 2013a). The second layer of soil of ISBA between 1 and 4 cm depth (the diffusion scheme is used in this study) is compared to in situ measurements at 5 cm depth at a 3-hourly time step (model output) between April and September in order to avoid frozen conditions as much as possible . The ability of ei_S, e5ei_S and e5_S to reproduce surface soil moisture variability is first assessed using the correlation coefficient (R) and unbiased root mean square differences (ubRMSD). Climatology differences between model and in situ observations make a direct comparison difficult (Koster et al., 2009b). Soil moisture time series usually show a strong seasonal pattern possibly increasing the skill values between modelled and observed data sets. To avoid seasonal effects, monthly anomaly time series are calculated. At each grid and observation point, the difference from the mean is produced for a sliding window of 5 weeks, and the difference is scaled to the standard deviation as in Albergel et al. (2013b). For each surface soil moisture estimate at day i, a period F is defined, with F=[i-17,i+17] (corresponding to a 5-week window). If at least five measurements are available in this period, the average soil moisture value and the standard deviation are calculated. Anomaly time series reflect the time-integrated impact of antecedent meteorological forcing. The latter is mainly reflected in the upper layer of soil. The correlation coefficient is also computed for anomaly time series (Rano). For correlations, the p value (a measure of the correlation significance) is also calculated indicating the significance of the test (as in Albergel et al., 2010), and only cases where the p value is below 0.05 (i.e. the correlation is not a coincidence) are retained. Stations with nonsignificant R values can be considered suspect and are excluded from the computation of the network average metrics. This process may remove some reliable stations too (e.g. in areas where the model might not realistically represent soil moisture).

Over the period 2010–2016, river discharge from ei_S, e5ei_S and e5_S are compared to daily streamflow data from the USGS http://nwis.waterdata.usgs.gov/nwis, last access: June 2018). Data are chosen for subbasins with large drainage areas (10 000 km2 or greater) and with a long observation time series (4 years or more). Smaller basins are excluded due to the low resolution of CTRIP (0.5∘ × 0.5∘). It is common to express observed and simulated river discharge (Q) data in m3 s-1. Given that the observed drainage areas may differ slightly from the simulated ones, specific discharge in mm d-1 (the ratio of Q to the drainage area) is used in this study, similarly to Albergel et al. (2017). Stations with drainage areas differing by more than 20 % from the simulated ones are also discarded. This criterion aims to ensure a meaningful comparison between observed and simulated values. It is necessary for coping with the significant distortions in the model representation of the river network that are caused by the coarse spatial resolution of the CTRIP global river network (0.5∘ × 0.5∘). Impact on Q is evaluated using the efficiency score (NSE; Nash and Sutcliffe, 1970). NSE evaluates the model ability to represent the monthly discharge dynamics and is given by NSE=1-∑t=1TQst-Qot2∑t=1TQot-Qot‾2, where Qst is the simulated river discharge (by either ei_S, e5ei_S or e5_S) at time t and Qot is observed river discharge at time t, T is the total number of days and Qot‾ is the average observed discharge. NSE can vary between -∞ and 1. A value of 1 corresponds to identical model predictions and observed data. A value of 0 implies that the model predictions have the same accuracy as the mean of the observed data. Negative values indicate that the observed mean is a more accurate predictor than the model simulation. Only stations with a NSE greater than -1 for the benchmark experiment, ei_S, are considered, leading to 172 stations over the considered domain. A normalized information contribution (NIC; as in Kumar et al., 2009) measure is then computed to quantify the improvement or degradation due to the specific atmospheric reanalysis used to force ISBA. The NICNSE values are computed for both e5_S and e5ei_S with respect to ei_S as NICNSE(e5;5ei)=NSE(e5;e5ei)-NSE(ei)1-NSE(ei). The NICNSE metric provides a normalized measure of the improvement through the use of either NSEe5ei or NSEe5 as a fraction of the maximum possible skill improvement (1-NSEei). Positive and negative NICNSE values indicate improvements and degradations in either e5_S or e5ei_S relative to ei_S river discharge estimates, respectively. NICs along with their 95 % confidence interval of the median derived from a 10 000 samples bootstrapping are provided for e5_S and e5ei_S. The ratio of simulated and observed river discharges is also computed Qst/Qot; the closer to 1 it is, the better the simulated river discharges are.

The Global Historical Climatology Network (GHCN) daily data set, developed to meet the needs of climate analysis and monitoring studies that require data at a daily time resolution, contains records from over 75 000 stations in 179 countries and territories (Menne et al., 2012a, b). Numerous daily variables are provided, including maximum and minimum temperature, total daily precipitation, snowfall and snow depth. In this study, over North America, stations with daily snow depth data from 2010 to 2016, with less than 10 % missing and at least 15 days of snow presences per year on average (to avoid using stations always reporting zero snow depth) are used, it results in 1901 stations out of 2056. The ability of ei_S, e5ei_S and e5_S to reproduce snow depth and its variability is assessed using the bias, correlation coefficient (R) and unbiased root mean square difference (ubRMSD).

Daily observations of sensible and latent heat fluxes from the FLUXNET-2015 data set with at least 2 years of data are used over the period 2010–2016 to evaluate the ability of e5_S, e5ei_S and ei_S to reproduce flux variability. The FLUXNET-2015 data set includes data collected at sites from multiple regional flux networks as well as several improvements to the data quality control protocols and the data processing pipeline (http://fluxnet.fluxdata.org/data/fluxnet2015-dataset/). The 37 stations are retained for the evaluations and two metrics are considered: R and RMSD.

Performance metrics are applied to each individual station of each network; thereafter, network metrics are computed by providing the median values of the statistics from the individual stations within each network. For each metric, the 95 % confidence interval of the median derived from a 10 000 samples bootstrapping is provided.

In response to the GCOS (Global Climate Observing System) endorsement of soil moisture as an essential climate variable, the European Space Agency Water Cycle Multimission Observation Strategy (WACMOS) project and Climate Change Initiative (CCI; http://www.esa-soilmoisture-cci.org, last access: June 2018) have supported the generation of a surface soil moisture product based on multiple microwave sources (ESA CCI SSM hereafter). The first version of the combined product was released in June 2012 by the Vienna University of Technology (Liu et al., 2011, 2012; Wagner et al., 2012). Several authors (e.g. Albergel et al., 2013a, b; Dorigo et al., 2015, 2017) have highlighted the quality and stability over time of the product. Despite some limitations, this data set has already shown potential in assessing model performance (e.g. Szczypta et al., 2014; van der Schrier et al., 2013). In this study the combined ESA CCI SSM latest version of the product (v4) is used. It merges SSM observations from seven microwave radiometers (SMMR, SSM/I, TMI, ASMR-E, WindSat, AMSR2, SMOS) and four scatterometers (ERS-1, 2 AMI, MetOp-A and B ASCAT) into a combined data set covering the period November 1978 to December 2016. Data are in volumetric (m3 m-3) units and quality flags (snow coverage, temperature below 0∘C or dense vegetation) are provided. For a more comprehensive overview of the product, see Dorigo et al. (2015, 2017). As topographic relief is known to negatively affect remote sensing estimates of soil moisture (Mätzler and Standley, 2000), the time series for pixels whose average altitude exceeded 1500 m above sea level were discarded. Data on pixels with urban land cover fractions larger than 15 % were also discarded, to limit the effects of artificial surfaces. The altitude and urban area thresholds were set according to Draper et al. (2011) and Barbu et al. (2014), who processed satellite-based SSM retrievals for data assimilation studies with the ISBA LSM. As for in situ measurements of soil moisture, correlation is applied to both the volumetric and anomaly time series.

The GEOV1 LAI used in this study is produced by the European CGLS (http://land.copernicus.eu/global/) as evaluated in Boussetta et al. (2015). The LAI observations are retrieved from the SPOT-VGT and then PROBA-V (from 1999 to present) satellite data according to the methodology proposed by Baret et al. (2013). As in Barbu et al. (2014), the 1 km spatial resolution observations are interpolated by an arithmetic average to the 0.5∘ model grid points, if at least 50 % of the observation grid points are observed (i.e. half the maximum amount). LAI observations have a temporal frequency of 10 days at best (in presence of clouds, no observations are available). Correlation and root mean squared differences are used to assess the ability of ei_S, e5ei_S and e5_S to reproduce LAI variability.

The GLEAM product uses a set of algorithms to estimate both terrestrial evaporation and RZSM based on satellite data (Miralles et al., 2011). It is a useful validation tool to assess model performance given that such quantities are difficult to measure directly on large scales. Potential evaporation rates are constrained by satellite-derived SSM data, while the global evaporation model in GLEAM is mainly driven by various microwave remote-sensing observations. It is now a well established data set that has been widely used to study land–atmosphere feedbacks (e.g. Miralles et al., 2014b; Guillod et al., 2015), as well as trends and spatial variability in the hydrological cycle (e.g. Jasechko et al., 2013; Greve et al., 2014; Miralles et al., 2014a; Zhang et al., 2016). This study makes use of the latest version available, v4.0. It is a 37-year data set spanning from 1980 to 2016 and is derived from a variety of sources, such as vegetation optical depth and snow water equivalent, satellite-derived SSM estimates, reanalysis air temperature and radiation, and a multi-source precipitation product (Martens et al., 2017). It is available at a spatial resolution of 0.25∘ × 0.25∘. A full description of the data set, including an extensive validation using measurements from 64 eddy-covariance towers worldwide is provided by Martens et al. (2017). As for LAI, the correlation and root mean squared differences are the two performance metrics used to evaluate the representation of evapotranspiration from the three data sets.

The final product used in this study is a daily GPP estimate from the FLUXCOM project (Jung et al., 2017). It is an upscaled product derived from the FLUXNET. In FLUXCOM, selected machine-learning-based regression tools that span the full range of commonly applied algorithms (from model tree ensembles to multiple adaptive regression splines, to artificial neural networks, and to kernel methods), and several representatives of each family are used to provide a spatial upscaling of GPP at regional to global scales. It is limited to a 0.5∘ × 0.5∘ spatial resolution and a daily temporal resolution over the period 1982–2013 (Tramontana et al., 2016). FLUXCOM fluxes can be used as a way of benchmarking LSMs on large scales (Jung et al., 2009, 2010, 2011; Beer et al., 2010; Bonan et al., 2011; Slevin et al., 2017). The product can be found at the Max Planck Institute for Biogeochemistry data portal (https://www.bgc-jena.mpg.de/geodb/projects/Home.php, last access: June 2018). Correlation and root mean squared differences are the two performance metrics used to evaluate the representation of carbon uptake from the three data sets.