Sequential assimilation of satellite-derived vegetation and soil moisture products using SURFEX_v8.0: LDAS-Monde assessment over the Euro-Mediterranean area

In this study, a global land data assimilation system (LDAS-Monde) is applied over Europe and the Mediterranean basin to increase monitoring accuracy for land surface variables. LDAS-Monde is able to ingest information from satellite-derived surface soil moisture (SSM) and leaf area index (LAI) observations to constrain the interactions between soil–biosphere–atmosphere (ISBA, Interactions between Soil, Biosphere and Atmosphere) land surface model (LSM) coupled with the CNRM (Centre National de Recherches Météorologiques) version of the Total Runoff Integrating Pathways (ISBA-CTRIP) continental hydrological system. It makes use of the CO2-responsive version of ISBA which models leaf-scale physiological processes and plant growth. Transfer of water and heat in the soil rely on a multilayer diffusion scheme. SSM and LAI observations are assimilated using a simplified extended Kalman filter (SEKF), which uses finite differences from perturbed simulations to generate flow dependence between the observations and the model control variables. The latter include LAI and seven layers of soil (from 1 to 100 cm depth). A sensitivity test of the Jacobians over 2000–2012 exhibits effects related to both depth and season. It also suggests that observations of both LAI and SSM have an impact on the different control variables. From the assimilation of SSM, the LDAS is more effective in modifying soil moisture (SM) from the top layers of soil, as model sensitivity to SSM decreases with depth and has almost no impact from 60 cm downwards. From the assimilation of LAI, a strong impact on LAI itself is found. The LAI assimilation impact is more pronounced in SM layers that contain the highest fraction of roots (from 10 to 60 cm). The assimilation is more efficient in summer and autumn than in winter and spring. Results shows that the LDAS works well constraining the model to the observations and that stronger corrections are applied to LAI than to SM. A comprehensive evaluation of the assimilation impact is conducted using (i) agricultural statistics over France, (ii) river discharge observations, (iii) satellite-derived estimates of land evapotranspiration from the Global Land Evaporation Amsterdam Model (GLEAM) project and (iv) spatially gridded observationbased estimates of upscaled gross primary production and evapotranspiration from the FLUXNET network. Comparisons with those four datasets highlight neutral to highly positive improvement.

Abstract. In this study, a global land data assimilation system (LDAS-Monde) is applied over Europe and the Mediterranean basin to increase monitoring accuracy for land surface variables. LDAS-Monde is able to ingest information from satellite-derived surface soil moisture (SSM) and leaf area index (LAI) observations to constrain the interactions between soil-biosphere-atmosphere (ISBA, Interactions between Soil, Biosphere and Atmosphere) land surface model (LSM) coupled with the CNRM (Centre National de Recherches Météorologiques) version of the Total Runoff Integrating Pathways (ISBA-CTRIP) continental hydrological system. It makes use of the CO 2 -responsive version of ISBA which models leaf-scale physiological processes and plant growth. Transfer of water and heat in the soil rely on a multilayer diffusion scheme. SSM and LAI observations are assimilated using a simplified extended Kalman filter (SEKF), which uses finite differences from perturbed simulations to generate flow dependence between the observations and the model control variables. The latter include LAI and seven layers of soil (from 1 to 100 cm depth). A sensitivity test of the Jacobians over 2000-2012 exhibits effects related to both depth and season. It also suggests that observations of both LAI and SSM have an impact on the different control variables. From the assimilation of SSM, the LDAS is more effective in modifying soil moisture (SM) from the top layers of soil, as model sensitivity to SSM decreases with depth and has almost no impact from 60 cm downwards. From the assimilation of LAI, a strong impact on LAI itself is found. The LAI assimilation impact is more pronounced in SM layers that contain the highest fraction of roots (from 10 to 60 cm). The assimilation is more efficient in summer and autumn than in winter and spring. Results shows that the LDAS works well constraining the model to the observations and that stronger corrections are applied to LAI than to SM. A comprehensive evaluation of the assimilation impact is conducted using (i) agricultural statistics over France, (ii) river discharge observations, (iii) satellite-derived estimates of land evapotranspiration from the Global Land Evaporation Amsterdam Model (GLEAM) project and (iv) spatially gridded observationbased estimates of upscaled gross primary production and evapotranspiration from the FLUXNET network. Comparisons with those four datasets highlight neutral to highly positive improvement. moisture (SSM and RZSM, respectively) and leaf area index (LAI) in a way that is fully consistent with the representation of surface and energy flux and river discharge simulations. Soil moisture (SM) is an essential component in partitioning incoming water and energy over land, thus affecting the variability of evapotranspiration, runoff and energy fluxes (Mohr et al., 2000). By controlling land surface temperature, plant water stress, evapotranspiration and infiltration of precipitation, soil moisture drives ecosystem dynamics, biodiversity and food production; regulates CO 2 emissions (uptake) from the land surface; and impacts natural hazards such as floods and droughts . The role of soil moisture as a regulator for various processes in the terrestrial ecosystem such as plant phenology, photosynthesis, biomass allocation and soil respiration, and hence the terrestrial carbon balance, has also clearly been established (Ciais et al., 2005;Van der Molen et al., 2012;Carvalhais et al., 2014;Reichstein et al., 2013). The seasonal dynamics of vegetation properties, such as LAI, are connected to soil moisture dynamics (Kochendorfer and Ramirez, 2010). Both the simulation of hydrological processes and the exchange of water vapour and CO 2 between the vegetation canopy and atmosphere interface are strongly influenced by LAI Szczypta et al., 2014).
Global observations of land surface variables are now operationally available from spaceborne instruments and they can be used to constrain LSMs through data assimilation (DA) techniques as demonstrated by several authors (e.g. Reichle et al., 2002;Draper et al., 2011Draper et al., , 2012Dharssi et al., 2011;Barbu et al., 2011;de Rosnay et al., 2013de Rosnay et al., , 2014Barbu et al., 2014;Boussetta et al., 2015;Fairbain et al., 2017). Recent studies (e.g. Traore et al., 2014) have demonstrated that a model that performs best for soil moisture does not necessarily perform best for plant productivity, highlighting the need to jointly use soil moisture and vegetation observations to improve global and continental eco-hydrological and/or carbon cycle models Kaminski et al., 2013). Several studies demonstrated the benefit of jointly assimilating SSM and LAI on the representation of RZSM (e.g. Sabater et al., 2008) and CO 2 flux (e.g. Albergel et al., 2010;Barbu et al., 2011).
Within the SURFEX modelling system (SURFace EXternalisée, Masson et al., 2013) the CO 2 -responsive version of the ISBA (Interaction Sol-Biosphère-Atmosphère: interaction between soil-biosphere-atmosphere) LSM (Noilhan and Mahfouf, 1996;Calvet et al., 1998Calvet et al., , 2004Gibelin et al., 2006) allows the representation of various land surface processes, including evapotranspiration and SM evolution. It is also capable of modelling photosynthesis and vegetation growth. The evolution of the simulated LAI and vegetation biomass changes in response to the meteorological forcing conditions. In previous studies, Barbu et al. (2014) and Fairbairn et al. (2017) tested a combined assimilation of SSM and LAI in this CO 2 -responsive version of ISBA over France within SURFEX. They used the force-restore version of ISBA (with three layers of soil), a simplified formulation of an extended Kalman filter (SEKF) with a 24 h assimilation window and hourly meteorological forcing from the SAFRAN reanalysis (Système d'Analyse Fournissant des Renseignements Atmosphériques à la Neige, Quintana-Seguıet al., 2008;Habets et al., 2008) on an 8 km scale. Fairbairn et al. (2017), also made a posterior offline use of runoff and drainage fields from ISBA to run the MODCOU hydrological model (MODèle COUplé, Habets et al., 2008) to evaluate the added value of the joint assimilation of LAI and SSM on the representation of river discharge over France. However, the assimilation was not successful in improving the representation of river discharge within MODCOU compared to an open-loop (i.e. no assimilation) simulation. Following their work, the present study tests the assimilation of both satellite-derived SSM and LAI on the continental scale. Further steps are made by the following: using the most recent SURFEX_v8.0 Offline Data Assimilation implementation; considering a much larger domain (Europe and the Mediterranean basin) as well as a longer time period, 2000-2012; using the multi-layer soil diffusion scheme of ISBA developed by Decharme et al. (2011).
Section 2 presents the LDAS-Monde system, i.e. (i) the CO 2responsive version of the ISBA LSM and the soil diffusion scheme, (ii) the CTRIP hydrological model and its coupling with ISBA, (iii) the atmospheric forcing used to drive the system, (iv) the equations of the SEKF, and (v) the assimilated remotely sensed observation dataset as well as the datasets used to assess the analysis impact. The latter is evaluated using agricultural statistics over France, river discharge, satellite-derived estimates of land transpiration and spatially gridded estimates of upscaled gross primary production from the FLUXNET network. Section 3 investigates and discusses the model sensitivity to the assimilated observations and provides a set of statistical diagnostics to assess and evaluate the analysis impact. Finally Sect. 4 provides perspective and future research directions.
2 Materials and method 2.1 SURFEX offline data assimilation The SURFEX modelling system includes the ISBA land surface model (Noilhan and Mahfouf, 1996) to calculate the soil, vegetation, snow energy and water budgets and is coupled to the TRIP (Oki et al., 1998) river routing model in order to simulate the streamflow (SURFEX-CTRIP hereafter). SURFEX offline data assimilation implementation is used to set up a land data assimilation system (LDAS) over Europe and the Mediterranean basin (longitudes from 11.75 • W to 62.50 • E, latitudes from 25.00 to 75.50 • N). It is defined as an offline sequential data assimilation system based on the ISBA LSM. It is capable of ingesting information from various satellite-derived observations to analyse and update SM and LAI simulated by ISBA. Analysis of ISBA prognostic variables then have an impact on the CTRIP variables (e.g. river discharge) through an interactive daily coupling (Voldoire et al., 2017). The system is driven by WFDEI observationbased atmospheric forcing dataset (Weedon et al., 2011(Weedon et al., , 2014. The main components of the LDAS (LSM, river routing system, analysis scheme and atmospheric forcing) are described in the following sections.

ISBA land surface model
ISBA models the basic land surface physics requiring only a small number of model parameters. The latter depend on the soil and vegetation types. This study uses the CO 2responsive version of ISBA, which is able to simulate the interaction between water and carbon cycles, photosynthesis, and its coupling to stomatal conductance (Calvet et al., 1998(Calvet et al., , 2004Gibelin et al., 2006). The CO 2 -responsive version of ISBA has been developed to allow for different biomass reservoirs for the simulation of photosynthesis and the vegetation growth. The dynamic evolution of the vegetation biomass and LAI variables is driven by photosynthesis in response to atmospheric and climate conditions. Photosynthesis enables vegetation growth resulting from the CO 2 uptake. During the growing phase, enhanced photosynthesis corresponds to a CO 2 uptake, which results in vegetation growth from the LAI minimum threshold (prescribed as 1 m 2 m −2 for coniferous forest or 0.3 m 2 m −2 for other vegetation types). In contrast, a deficit of photosynthesis leads to higher mortality rates. The total evaporative flux represents the combination of the 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 the snow and soil ice. The CO 2 uptake from photosynthe-sis is defined as the gross primary production (GPP) and the release of CO 2 is called the ecosystem respiration (RECO). The net ecosystem CO 2 exchange (NEE) measures the difference between these two quantities. ISBA has an explicit snow scheme (with 12 layers), as detailed in Bonne and Etchevers (2001) and Decharme et al. (2016). The multi-layer soil diffusion scheme version (ISBA-Dif) is based on the mixed form of the Richards equation (Richards, 1931) and explicitly solves the one-dimensional Fourier law. Additionally, ISBA-Dif incorporates soil freezing processes developed by Boone et al. (2000) and Decharme et al. (2013). The total soil profile is vertically discretized and the temperature and the moisture of each layer are computed according to the 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 . A discretization with 14 layers over 12 m depth is used. The lower boundary of each layer is 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). The amounts of clay, sand and organic carbon present in the soil are determined by thermal and hydrodynamic soil properties (Decharme et al., 2016) and are taken from the Harmonised 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 discrepancy between the surface runoff and the throughfall rate is defined by the infiltration rate.
The throughfall rate is defined as the sum of rainfall that is not intercepted by the canopy, dripping from the canopy (interception reservoir), and snow melt water. Evaporation only affects the superficial layer, which represents the top 1 cm of soil. The soil evaporation is proportional to the relative humidity of the superficial layer. Transpiration water from the root zone (the region where the roots are asymptotically distributed) follows the equations in Jackson et al. (1996). More information on the root density profile is available in Canal et al. (2014). ISBA total runoff has two contributions: the surface runoff (the lateral subsurface flow in the topsoil) and a free drainage condition at the bottom layer. A basic TOPMODEL approach is used to compute the Dunne runoff (i.e. when no further soil moisture storage is available) and lateral subsurface flow from a sub-grid distribution of the topography. The Horton runoff (i.e. when rainfall has exceeded infiltration capacity) is estimated from the maximum soil infiltration capacity and a sub-grid exponential distribution of the rainfall intensity.

CTRIP river routing
The present CTRIP version consists of a global streamflow network at 0.5 • spatial resolution. The CTRIP model is driven by the three prognostic equations corresponding to the groundwater, the surface stream water and the seasonal floodplains. Streamflow velocity is computed using the Manning formula (Decharme et al., 2010). The floodplain reservoir fills when the river water level overtops the riverbank and empties again when the water level drops below this threshold . Flooding impacts the ISBA soil hydrology through infiltration. It also influences the overlying atmosphere via free surface water evaporation and precipitation interception.
At last, the groundwater scheme (Vergnes and Decharme, 2012) is based on the two-dimensional groundwater flow equation for the piezometric head. Its coupling with ISBA permits the presence of a water table under the soil moisture column to be accounted for, allowing 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 returns to ISBA the water table depth or rise, floodplain fraction, and flood potential infiltration.

Extended Kalman filter
This section describes the analysis update of the extended Kalman filter while its application setup is described in Sect. 2.3.
The analysis update equation of the extended Kalman filter is as follows: (1) The "a", "f" and "o" subscripts stand for analysis, forecast and observation, respectively. The term x denotes the control vector of dimension N x , computed at time t i , that represents the prognostic equations of the LSM M.
The term y o denotes the observation vector of dimension N y . The Kalman gain matrix K i is computed at time t i as follows: A non-linear observation operator h enables the extraction of the model counterpart of the observations: B and R are error covariance matrices characterizing the forecast and observation vectors. The cross-correlated terms represent covariances. The operator H (and its transpose H T ) from Eq. (2) is the Jacobian matrix: the linearized version of the observation operator (defined as N y rows and N x columns) that transforms the model states into the observation space. A numerical estimation of each Jacobian element is calculated by finite differences, by perturbing each component x j of the control vector x by a specific amount δx j , resulting in a column of the matrix H for each integration m: The control vector evolution from time t i to the end of the 24 h assimilation window (t i+1 ) is then controlled by the following equation: In line with previous studies (e.g, Mahfouf et al., 2009;Albergel et al., 2010;Barbu et al., 2011Barbu et al., , 2014de Rosnay et al., 2013;Fairbairn et al., 2015Fairbairn et al., , 2017 a fixed estimate of the background-error variances and zero covariances at the start of each cycle are used, leading to a SEKF. The initial state at the start of a 24 h assimilation window is analysed by assimilating the observations available over the previous 24 h assimilation window. This approach is similar to the "simplified 2-D-Var (two-dimensional variational data assimilation scheme)" proposed by Balsamo et al. (2004), but the increments are applied at the final timestep of the 24 h assimilation window. Draper et al. (2009) found that the SEKF could generate flow dependence from the 24 h assimilation window and that cycling the background-error covariance (as in the EKF) gave no additional benefit.

WFDEI observation-based atmospheric forcing dataset
Atmospheric forcing from the WFDEI dataset (Weedon et al., 2011(Weedon et al., , 2014) is used to drive the LDAS. It spans the period 1979-2012 and contains 3-hourly time intervals of wind speed, atmospheric pressure, air temperature (averaged values are used), air humidity, incoming shortwave and longwave radiations, and solid and liquid precipitation. WFDEI originates from the ECMWF ERA-Interim reanalysis (Dee et al., 2011), interpolated to a spatial resolution of 0.5 • , and is corrected with the CRU dataset (Climatic Research Unit, Harris et al., 2014) using a sequential elevation correction of surface meteorological variables plus monthly bias correction from gridded observations (e.g. precipitation data from the Global Precipitation Climatology Centre; GPCC). A more exhaustive description of the dataset is available in Schellekens et al. (2017).

ESA CCI surface soil moisture
This study makes use of a multi-sensor, long-term and global satellite-derived surface soil moisture dataset (Liu et al., 2011Wagner et al., 2012;Dorigo et al., 2015 developed within The European Space Agency Water Cycle Multi-mission Observation Strategy (ESA-WACMOS) project and Climate Change Initiative (CCI, http://www.esa-soilmoisture-cci.org). Several authors (e.g. Albergel et al., 2013a, b;Dorigo et al., 2015) have highlighted the quality and stability over time of the product. Despite some limitations, this dataset has shown potential for assessing model performance (Szczypta et al., 2014; To assimilate SM data, it is important to rescale the observations such that they are consistent with the model climatology (Reichle and Koster, 2004;Drusch et al., 2005). The climatology of the SM dataset is defined by the specific mean value, variability and dynamical range. The ISBA model climatology for each grid point is dependent on the dynamical range, which is calculated from the wilting point and field capacity parameters (functions of soil texture types). It is necessary to transform the ESA CCI SSM product into model-equivalent SSM to address possible mis-specification of physiographic parameters, such as the wilting point and the field capacity. The linear rescaling approach described in Scipal et al. (2008) (using the first two moments of the cumulative distribution function, CDF) has been used in this study; it is a linear rescaling that enables a correction of the differences in the mean and variance of the distribution. The first two moments, the intercept a and the slope b, are as follows: where SSM m , SSM o , σ m and σ o correspond to the model and observation means and standard deviations (SDs), respectively. Barbu et al. (2014) and Draper et al. (2011) discussed the importance of allowing for seasonal variability in the CDF matching. Parameters a and b vary spatially and were derived on a monthly basis by using a 3-month moving window over 2000-2012 after screening for the presence of ice and urban areas. The ESA CCI SSM observations are interpolated by an arithmetic average to the 0.5 • model grid points.

GEOV1 leaf area index
The GEOV1 LAI is produced by the European Copernicus Global Land Service project (http://land.copernicus.eu/ global/). The LAI observations are retrieved from the SPOT-VGT and PROBA-V (from 1999 to present) satellite data according to the methodology discussed in Baret et al. (2013). Following Barbu et al. (2014), the 1 km resolution observations are interpolated by an arithmetic average to the 0.5 • model grid points, as long as 50 % of the observation grid points are observed (half the maximum amount). LAI observations have a temporal frequency of 10 days. Both SSM and LAI observed datasets are illustrated in Fig. 1, which presents averaged values over 2000-2012. Figure 1 also illustrates the studied domain.

Evaluation datasets and strategies
A common diagnostic in data assimilation is to compute Eq. (1) differences between the assimilated observations and the model background, called the innovations, and Eq. (2) differences between the assimilated observations and the analysis, called the residuals (Barbu et al., 2011). Assuming that the system is working well, residuals have to be reduced compared to the innovations. After evaluating innovations and residuals of SSM and LAI, analysis impact is assessed using Eq. (1) agricultural statistics over France, Eq. (2) observed river discharge over Europe, Eq. (3) satellite-derived estimates of terrestrial evapotranspiration from the Global Land Evaporation Amsterdam Model (GLEAM, Martens et al., 2017) and Eq. (4) spatially gridded estimates of upscaled GPP and evapotranspiration from the FLUXNET network (Jung et al., 2009, 2011). Smith et al. (2010a demonstrated that crop simulations can be validated by agricultural statistics on a country scale. With a finer spatial scale over France, Calvet et al. (2012) benchmarked several configurations of the ISBA LSM using agricultural statistics (Agreste, 2016), namely the correlation between yield time series and above-ground biomass (B ag ) simulations. In ISBA, B ag of herbaceous vegetation is made up of two components: the active biomass and the structural biomass. The former describes the photosynthetically active leaves and is linked to B ag by a nitrogen dilution allometric logarithmic law (Calvet and Soussana, 2001). Calvet et al. (2012), found that B ag simulated by the model is in agreement with the agricultural statistics, and therefore can be used to benchmark model and system development. Yearly statistical surveys over France are provided by the Agreste portal (http://agreste.agriculture.gouv.fr/). This has enabled a database of annual straw-cereal grain yield (GY) values to be established. The GY estimates are available according to administrative unit (département) and per crop type. Following Calvet et al. (2012), Canal et al. (2014) and Dewaele et al. (2017), the GY values for rainfed straw cereals over 45 départements are used, which include barley, oat, rye, triticale and wheat. Simulated and analysed annual maximum of B ag are compared to GY estimates following the methodology from Dewaele et al. (2017). Although SURFEX does not directly represent GY, it is assumed that the regional-scale simulations of above-ground biomass from a generic LSM can provide the inter-annual variability as a proxy for GY (Calvet et al., 2012;Canal et al., 2014).
Over 2000-2010, simulated and analysed river discharge is compared to gauging measurements from the Global Runoff Data Center (GRDC; http://grdc.sr.unh.edu/index. html) and the Banque Hydro (http://www.hydro.eaufrance. fr/) at a monthly time step. Data are chosen over the domain presented in Fig. 1 for sub-basins with large drainage areas (10 000 km 2 or greater) and with a long observation time series (4 years or more). It is common to express observed and simulated river discharge (Q) data in cubic metres per second (m 3 s −1 ). However, given that the observed drainage areas may differ slightly from the simulated ones, scaled Q values in millimetres per day (mm d −1 , the ratio of Q to the drainage area) are used in this study. Stations with drainage areas differing by more than 15 % from the simulated (analysed) ones are also discarded. This leads to 83 stations. Impact on Q is evaluated using correlation, root mean square difference (RMSD) and the efficiency score (Eff) (Nash and Sutcliffe, 1970). Eff evaluates the model's ability to represent the monthly discharge dynamics and is given by the following: where Q t s is the simulated river discharge (or analysed) at time t, and Q t o is observed river discharge at month mt. The Eff 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.
The GLEAM product uses a set of algorithms to estimate terrestrial evaporation and root-zone SM from satellite data (Miralles et al., 2011). It is a useful validation tool given that such quantities are difficult to measure directly on large scales. The global evaporation model in GLEAM is mainly driven by microwave remote sensing observations, while potential evaporation rates are constrained by satellite-derived SM data. It is a well-established dataset that has been widely used to study 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) and land-atmosphere feedbacks (e.g. Miralles et al., 2014b;Guillod et al., 2015). This study makes use of the latest version available, v3.0. It is a 35-year dataset spanning from 1980 to 2014 and is derived from a variety of sources, namely vegetation optical depth (VOD) and snow water equivalents (SWEs), satellite-derived SM, reanalysis air temperature and radiation, and a multisource precipitation product (Martens et al., 2017). It is available at a spatial resolution of 0.25 • . Martens et al. (2017) provide a full description of the dataset, including an extensive validation using measurements from 64 eddy-covariance towers worldwide.
The upscaled FLUXNET GPP and evapotranspiration were derived from the FLUXNET network using a model tree ensemble (FLUXNET-MTE hereafter) approach as described in Jung et al. (2009). It is a machine learning technique that can be trained to ascertain land-atmosphere fluxes, providing a way of benchmarking LSMs on large scales (Jung et al., 2009Beer et al., 2010;Bonan et al., 2011;Jung et al., 2011;Slevin et al., 2017). The machine learning algorithm is trained using a combination of land cover data, observed meteorological data and remotely sensed vegetation properties (fraction of absorbed photosynthetic active radiation). The algorithm uses model tree ensembles to provide estimates of carbon fluxes at FLUXNET sites with available quality-filtered flux data, after which the trained model can be implemented globally using grids of the input data (Jung et al., 2009(Jung et al., , 2011. It is limited to a 0.5 • spatial resolution and a monthly temporal resolution over a 20-year period . It can be found in the Max Planck Institute for Biogeochemistry Data Portal (https://www.bgc-jena.mpg.de/ geodb/projects/Home.php).

Experimental setup
The LDAS used in this study is designed as follows: x is the eight-dimensional control vector including soil layers 2 to 8 (representing a depth from 1 to 100 cm) and LAI propagated by the ISBA LSM. The term y o denotes the two-dimensional observation vector (SSM, LAI). The model counterparts of the observations are the second layer of soil of ISBA LSM (w 2 between 1 and 4 cm) and LAI for SSM and LAI observations, respectively. A comparison between ESA CCI SM and the two top ISBA soil layers suggests that the second layer of soil better represents the satellite-derived product (not shown). Also, the first layer of soil (1 cm depth) is discarded from the control vector as, over a 24 h window, it is more reactive to the atmospheric forcing than to a small initial perturbation (Draper et al., 2011;Barbu et al., 2014). This leads to the following expression of the Jacobians matrices: Several studies (e.g. Draper et al., 2009;Rüdiger et al., 2010) have demonstrated that small perturbations (10 −3 or less) lead to a good approximation of this linear behaviour, provided that computational round-off error is not significant. Following Draper et al. (2011) and Mahfouf et al. (2009), the soil moisture errors are assumed to be proportional to the dynamic range (the difference between the volumetric field capacity, w fc , and the wilting point, w wilt ), which is determined by the soil texture (Noilhan and Mahfouf, 1996); in this study the Jacobian perturbations were assigned values of 1×10 −4 ×(w fc − w wilt ). Following Rüdiger et al. (2010), the LAI perturbation was set to a fraction (0.001) of the LAI itself. In this configuration, for every 24 h analysis cycle, the LSM is run several times: first to get the model trajectory (forecast), then perturbing the initial conditions of the various control variables, allowing computation of the various terms of the Jacobians (Eq. 4).
For soil moisture in the second layer of soil, i.e. the model equivalent of the SSM observations, a mean volumetric SD error of 0.04 m 3 m −3 is prescribed. A smaller mean volumetric SD error of 0.02 m 3 m −3 is prescribed to the deeper layers, as suggested by several authors for RZSM Draper et al., 2011;Barbu et al., 2011Barbu et al., , 2014. The observational SSM error is set to 0.05 m 3 m −3 as in Barbu et al. (2014). This value is consistent with errors estimated from a range of remotely sensed soil moisture sources (e.g. de Jeu et al., 2008;Draper et al., 2011;Gruber et al., 2016). Soil moisture observational and background errors are also scaled by the model soil moisture range. The error SDs in the GEOV1 LAI and the modelled LAI (for modelled LAI values higher than 2 m 2 m −2 ) are both assumed to be equal to 20 % of the LAI values. In accordance with a study by Barbu et al. (2011), the modelled LAI values lower than 2 m 2 m −2 are assigned a constant error of 0.4 m 2 m −2 .
SURFEX-CTRIP was spun up by cycling 20 times through the year 1990, and then a 10-year model run is allowed before considering both an open loop (a model run with no assimilation) and an analysis experiment over [2000][2001][2002][2003][2004][2005][2006][2007][2008][2009][2010][2011][2012]. Diagnostic studies of the Jacobian values have usually been performed before, including new observation types Fillion and Mahfouf, 2003;Garand et al., 2001;Rudiger et al., 2010). That is why, following Rudiger et al. (2010), an analysis experiment without assimilating any observations has also been run over 2000-2012 to study the model sensitivity to the observations through the Jacobians. Table 1 summarizes the SURFEX-CTRIP setup used in this study.

Consistency between the model and observations
Consistency over time is crucial when assimilating long-term datasets. Several authors assessed the consistency of the ESA CCI soil moisture product with respect to re-analysis products (e.g. Loew et al., 2013;Albergel et al., 2013a, b) and in situ measurements (Dorigo et al., 2015. Lambin et al. (1999) found that the GEOV1 LAI dataset is also consistent over time and can be used, for example, for detection of change and for providing information on shifting trends or trajectories in land use and cover change. To verify the results from literature for the spatial and temporal domain considered in this study, a consistency evaluation both for SSM and LAI against the open-loop experiment has been performed. As observed SSM climatology is matched to the model climatology (see Sect. 2.2.2), consistency between observations and the model over time (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012), is expressed as correlations on both absolute and anomaly time series. The latter is computed using monthly sliding windows as described in Albergel et al. (2009). Only significant correlation values (at p value < 0.005) are retained. For LAI, consistency is expressed both as correlations and RMSDs.
Median soil moisture correlation (anomaly correlation), of ESA CCI SSM with the SURFEX-CTRIP second layer of soil, w 2 between 1 and 4 cm, is 0.65 (0.47) over 2000-2012. Year-to-year correlation (anomaly correlation), which can potentially be impacted by the prevailing conditions in the given years, ranges from 0.62 (0.45) to 0.71 (0.48). Although many different sensors are used over time and space to retrieve ESA CCI SSM, the product can be considered stable. Over the same period, correlation and RMSDs between GEOV1 LAI and SURFEX-CTRIP is 0.  served in winter time are not desirable since the vegetation is supposed to be dormant.
Overall both ESA CCI SSM and GEOV1 LAI were found to be stable over time with respect to SURFEX-CTRIP, as illustrated in Fig. 3 for 2000, 2006 and 2012. Figure 3a-c illustrate correlations between ESA CCI SSM and SURFEX-CTRIP (w 2 ). While in 2000 not all of Europe is covered, it is the case from 2003 onwards. Low correlation values are found in desert areas (over the Sahara), high elevation (e.g. over the Alps) and at high latitudes, whereas high correlation values are obtained over, for example, the Iberian Peninsula, France and Turkey. Figure 3d-f present the correlations and RMSD values for GEOV1 LAI with SURFEX-CTRIP, respectively, only for vegetated grid points (> 90 %). Generally, LAI at high elevation is not represented well (low correlations and high RMSD) as well as in the northeastern part of the domain, which is mainly covered by broad-leaved trees. Conversely, the southern part of the domain presents high levels of correlations and low RMSD values.

Model sensitivity to observations
The Jacobians, H (Eq. 4), are dependent on the model physics. The examination of the Jacobians provides useful insight into explaining the data assimilation system performances (Barbu et al., 2011;Fairbairn et al., 2017). Median values over 2000-2012 are presented in Table 2.
The model equivalent of SSM is the second layer of soil (w 2 between 1 and 4 cm depth). It is then expected that the sensitivity of SSM to changes in soil moisture of that layer is higher than those of the other layers of soil. Sensitivity of LAI to changes in soil moisture (Table 2, bottom rows) suggests that control variables related to soil moisture will also be impacted by the assimilation of LAI. The model sensitivity to SSM decreases with depth, as presented in Table 2 revealing that the assimilation of SSM will be more effective in modifying soil moisture from the first layers. Over Europe, median values of H with respect to SSM observations (Table 2 top rows) range from 0.1719 to 0.0001 for  layers w 2 to w 8 , respectively, and is -0.0001 for LAI. The negative value of ∂SSM t ∂LAI 0 also indicates that a positive increment of LAI will generally lead to a decrease of SSM (w 2 ). The depth impact is also illustrated in Fig. 4 which represents histograms of H over Europe for three control variables (w 2 in red, w 4 in cyan and w 8 in blue) with respect to a change in SSM for 6 months (January, March, June, August, October, December) over 2000-2012 ( Fig. 4a-f). Additionally Fig. 4 depicts a seasonal dependency. For instance, the histogram representing H of control variable w 2 (Fig. 4a) presents mainly three types, Eq. (1) values close or equal to 0 (type_A), Eq. (2) values between 0.2 and 0.8 (type_B), and Eq. (3) close to 1 (type_C). The values of type_C correspond to the situation in which the model dynamic is close to the identity, i.e. the perturbation of the initial state is almost unchanged by the end of the assimilation window (24 h). For values of type_B, the model dynamic is strongly dissipative and therefore the final offset is only a fraction of the initial perturbation. Distributions of types A, B and C vary in time; for January they are 75, 14 and 11 %, for June they are 36, 44 and 20 %, and for October they are 48, 30 and 22 %, respectively. It suggests a higher sensitivity of the first layers of soil to a change in SSM, particularly during late summer and autumn rather than during winter months. While a similar behaviour is observed up to the fourth layer of soil, the deepest layers of soil (e.g. w 8 , blue line) do not show any seasonal dependency, and very small sensitivity with mainly Jacobian values of type A.
The same typology can apply to H values ∂LAI t ∂LAI 0 (Fig. 4gi), with an even stronger seasonal dependency. For all months of January, distributions are 81, 18 and 1 %, while they are 22, 77 and 1 % for all months of June and 27, 45 and 28 % for all months of October for types A, B and C, respectively. Assimilation of LAI will be more effective in modifying LAI from late spring to autumn. Finally, the assimilation of LAI will be more effective in modifying soil moisture from layers 4 to 6 (Table 2).

Impact of the analysis on control variables
Control variables are directly impacted by the assimilation of LAI and SSM. Figure 5 illustrates averaged analysis increments for the period 2000-2012 for LAI and soil moisture in w 2 (between 1 and 4 cm), w 4 (between 10 and 20 cm) and w 6 (between 40 and 60 cm) for all months of February, May, August and October. Red (blue) colours indicate that the analysis removes (adds) LAI and soil moisture. At the beginning of the year vegetation is not very active, but on the very western part of the domain the analysis tends to add LAI over the United Kingdom and northwestern parts of France and it reduces LAI over the Iberian Peninsula. At the beginning of the year soil moisture is only slightly affected by the analysis. Later in spring and summer the analysis is more efficient: it removes LAI over a large part of Europe, reducing the bias observed between open loop and observations. It mainly adds water in w 2 and remove water from layers w 4 to w 6 . The seasonally marked impact of the analysis is consistent with the above description of the Jacobians behaviour. Analysis increments are also presented in Fig. 6 for the entire period 2000-2012. Generally, the analysis tends to remove LAI and add water in w 2 but dries layers where the roots are mainly located (from w 4 to w 6 ). Its effect is, however, less pronounced at greater depths. Figure 7 shows the averaged analysis impact on LAI for all months of January, April, July and October over 2000-2012 expressed in RMSD in the following way: GEOV1 LAI vs. open loop and GEOV1 LAI vs. analysis. Only points where observed LAI is available (and assimilated) are retained. As this impact assessment is conducted against the observations that were assimilated, improvements from the analysis are expected and show that the LDAS is working well. From  Fig. 7, this is mostly the case (e.g. in October). As indicated in Sect. 3.2, the analysis is most efficient during late summer and autumn. The geographical patterns highlighted in Sect. 3.1 are also observed with a clear improvement, e.g. in    the northeastern part of the domain. Analysis improvement with respect to the observations is also visible in Fig. 7. Figure 8 illustrates histograms of innovations (in red) and residuals (in green) of LAI for all months of February, April, June, August, October and December over 2000-2012. As expected, the distribution of residuals is more centred on 0 than the distribution of the innovations. A seasonal pattern can be observed: during the growing phase (and up to June), both innovations and residuals present a right tail indicating that the model (and the analysis to a lesser extent) tends to underestimate LAI. In this period, similarities between innovations and residuals suggest that the analysis is not very efficient. At the end of summer and in autumn, distributions present a left-tail distribution; LAI is overestimated but this time the analysis is more efficient. Distributions of SSM residuals are even more centred on zero than those of innovations with no seasonal dependency and smaller differences (not shown). The common CDF-matching technique applied to SSM to remove systematic errors is responsible for this smaller impact, as the LDAS can only correct SSM short-term variability. Contrary to SSM, the LAI mismatch between the open loop and the GEOV1 estimates concerns both magnitude and timing (see for example Fig. 6 in Barbu et al., 2014). face variables that are indirectly impacted by the assimilation, namely daily cumulated soil drainage flux, supersaturation runoff, evapotranspiration and daily mean river discharge. Although the analysis impact is relatively weak on those variables (e.g. ∼ 1 % on the river discharge represented over the Danube), geographical patterns are observed. Areas where positive analysis increments were found for LAI (Fig. 5) tend to correspond to a decrease in drainage and runoff (in red in Fig. 9) while evapotranspiration increases (in blue Fig. 9). Changes in these indirectly impacted land surface variables are in agreement with the analysis increment maps (Fig. 6).

Evaluation of analysis impact
First, the evaluation of the analysis impact is effectuated over France using straw cereal GY values from the Agreste French agricultural statistics portal. Only the département administrative units corresponding to a high proportion of straw cereals are considered. Yearly maximal above-ground biomass (B ag ) values from the open loop (i.e model) and analysis are compared to GY over 2000-2010. Yearly-scaled anomalies from the mean and the SD for observations, open loop and analysis are used for 45 sites over France as in Dewaele et al. (2017). Figure 10a and 10b present correlations and RMSD values, respectively, and Fig. 10c shows time series for one site illustrating the inter-annual variability. After assimilation of SSM and LAI, correlation as well as RMSD between B ag and GY is clearly improved for 43 and 41 sites, respectively, out of 45 sites showing the added value of the analysis compared to the open loop. Figure 10c    Over 2000-2010, 48 of 83 gauge stations present Eff values greater than 0, and 22 gauge stations report Eff greater than 0.5. As suggested in the previous section, the analysis impact on river discharge is rather small. If the analysis generally leads to an improvement in river discharge representation, only 8 stations present an Eff increase greater than to 0.05 (3 stations report a decrease greater than 0.05). Eff, R and RMSD histograms of differences are presented in Fig. 11b, c and d along with a hydro-graph (Fig. 11a) for the Loire River in France (47.25 • N, 1.52 • W). Although the assimilation impact is relatively small, evaluation results suggest that they are neutral to positive. Analysis impact on other CTRIP variables (e.g. floodplain fraction and storage, groundwater height) is rather neutral.
Evapotranspiration from both the open loop and the analysis are compared to monthly values of GLEAM satellitederived estimates over 2000-2012 for vegetated grid points (> 90 %). As for the river discharge, the assimilation impact on evapotranspiration is small. However, the comparison with the GLEAM satellite-derived estimates is rather positive, as illustrated in Fig. 12, which represents evapotranspiration from the open loop (Fig. 12a), GLEAM estimates (Fig. 12b), the analysis (Fig. 12c) and their differences (Fig. 12d). Open-loop simulation of evapotranspiration tends to over-estimate the GLEAM product over most of Europe, particularly over France, the Iberian Peninsula and northern Africa. Analysis is able to reduce this bias (Fig. 12d). Figure 14 shows maps of RMSD (Fig. 14a) and correlation (Fig. 14b) (Fig. 13a), FLUXNET-MTE (Fig. 13b) and the analysis (Fig. 13c) as well as differences between the analysis and the model (Fig. 13d). Also, Fig. 14c Evensen, 1994), to date SEKF is the most mature technique developed for land surface data assimilation within SUR-FEX. Many studies using SURFEX exposed the strengths and weaknesses of this technique Albergel et al., 2010., Draper et al., 2011Barbu et al., 2011Barbu et al., , 2014Duerinckx et al., 2015;Fairbairn et al., 2015Fairbairn et al., , 2017. The SEKF relies on accurate linear assumptions in deriving the Jacobians. Draper et al. (2009), Duerinckx et al. (2015 and Fairbairn et al. (2015) pointed out that outliers in Jacobian's values may occur under specific conditions (e.g. close to threshold values such as the wilting point and field capacity for soil moisture), possibly leading to instabilities in the analysis. Those outliers in the Jacobian's values were, however, obtained using the force-restore version of the ISBA LSM with three layers of soil and not with the diffusion soil scheme: ISBA-Dif. In such configurations they used only one control variable related to soil moisture; the second layer of soil was a thick layer representing all the rootzone (w 2−RZ ) while the model equivalent was the very thin top layer (∼ 1 cm). Thus ∂SSM t ∂w 2−RZ 0 Jacobians, representing the impact of perturbing w 2 (i.e. the whole root-zone) on SSM (∼ 1 cm) can be quite different compared to those obtained using the soil diffusion scheme and presented in this study (e.g. where w 2 and SSM representing the same depth; 1-4 cm). For instance, ∂SSM t ∂w 2−RZ 0 Jacobians exhibit a rather large proportion of negative values, as illustrated by Fig. 10 of Fairbairn et al. (2017) and discussed in Parrens et al. (2014).
Very few negative Jacobian values are obtained with the diffusion soil scheme (as in Fig. 4) over Europe for 2000-2012. The SEKF is also limited in correcting errors from the atmospheric forcing uncertainty making the model too reliant on the chosen forcing. Alternatively an EnKF, which relies on the ensemble spread to capture background errors, can be modified to stochastically capture both model and precipitation errors (Maggioni et al., 2012;Carrera et al., 2015). The use of an EnKF within LDAS-Monde is currently under investigation at Météo-France. Alternatively, particle filters could provide a means to capture non-Gaussian errors (e.g. Moradkhani et al., 2012). The performance of an analysis scheme depends on appropriate statistics for background and observation errors. Wrongly specified error parameterization may negatively affect the analysis. The main objective of this study was to present the newly developed LDAS-Monde while the statistics for background and observation errors were obtained from the literature. Soil moisture observations and background errors were scaled using the open-loop soil moisture dynamical range. This accounts for texture-based spatial variability in the error and assumes that the soil moisture errors and the dynamic range have a linear relationship. Time correlations in the errors have also been neglected in this study, which are likely to occur in reality. It is also possible to employ an a posteriori diagnostic to estimate observation errors, such as the statistics of the innovations (observations minus background) (Andersson, 2003;Mahfouf et al., 2007). This approach has been successfully applied by Barbu et al. (2011) in a point-scale experiment to obtain seasonal error variability; the approach does not provide objective estimates of the observational errors but assesses the sub-optimality of the analysis. Future work will investigate having spatially and temporally variant observations errors, based on statistical methods already applied to the ESA CCI SSM dataset such as triple collocation (Dorigo et al., 2015) or error decomposition .
Having LAI estimates every 10 days while using 24 h assimilation window may also trigger analysis discrepancies, as between two LAI assimilations the system relies only on SSM assimilation. When a large analysis update occurs on LAI (from the assimilation of LAI), it then tends to go back towards the model states in the successive days before being constrained again by the next observations. For instance, in winter most of the ∂LAI t ∂LAI 0 Jacobians are equal (or close) to zero and therefore the analysis update caused the LAI to return almost instantaneously to the incorrect LAI minimum value. This issue could be addressed using longer assimilation windows, from 10 days up to 1 month (e.g. as in Jarlan et al., 2008) where different data assimilation techniques could be used (e.g. variational methods to obtain a best fit between several observations). An alternative could be to keep a 1-day assimilation window and use smoothing techniques (e.g. Munier et al., 2014) to keep the benefit of the analysis update by propagating the error covariance forward up to the next available observation.
4.2 Can better use of the observations improve the analysis?
4.2.1 Towards a better use of GEOV1 leaf area index SURFEX_v8.0 does not use any crop-specific parameterization, which would be required to simulate the crop grain yield formation. In addition, the simulations of photosynthesis and vegetation growth do not take into account certain factors impacting the long-term agricultural production (e.g. changes in agricultural practices, diseases, pests, crop migration, the grain formation and crop cultivars). However, previous studies (Calvet et al., 2012;Canal et al., 2014) showed that agricultural statistics such as grain yields can be used to benchmark SURFEX development in representing the above-ground biomass inter-annual variability. A strong positive impact from the assimilation of SSM and LAI on the representation of above-ground biomass inter-annual variability has been highlighted in this study. The impact on river discharge representation is only small (Sect. 3.3). Improvements are, however, expected from a better representation in the model of vegetation parameters such as LAI (e.g. Szczypta et al., 2014). Although the analysis is efficient in correcting LAI, high RMSD values remain, particularly during the senescence phase when SURFEX-CTRIP over-estimates LAI over a large part of Europe. RMSD and correlations with GEOV1 and SURFEX-CTRIP still expose a strong seasonal dependency after the analysis (red line in Fig. 2) which is mainly attributed to model errors. The GEOV1 estimates have been shown to exhibit some realistic environmental features that are not, or are poorly, simulated by the model (Fairbairn et al., 2017). Therefore, it was decided not to rescale the GEOV1 estimates to the model climatology. Figure 2 also suggests that the minimum LAI values used as model parameters (see Sect. 2.1.1) should be revisited because such large differences are not desirable, particularly when the vegetation is dormant. Another caveat of this study is the use of a single LAI value for all vegetation types that are represented in SURFEX-CTRIP. As detailed in Barbu et al. (2014), the Kalman gain is calculated for each individual vegetation type (patch). The analysis increment is added to the background for each patch, producing a patch-dependent analysis update. The patch dependence is introduced in the Kalman gain via the Jacobian elements. The possibility of having LAI estimates for each type of vegetation is under investigation and has the capacity of overcoming the two above-mentioned weaknesses. Re-cently, the GEOV1 LAI data have been disaggregated following a Kalman filtering technique developed by Carrer et al. (2014). This enables the LAI signal for each patch to be separated within the pixel, which provides a dynamic patchdependent estimate of the assimilated LAI within the pixel . From the individual estimates over 1999-2015, minimum LAI values have also been used to update the model parameterization. Preliminary results from assimilating disaggregated LAI time series and using new LAI minimum values (not shown) suggest better representation of vegetation variables such as LAI and above-ground biomass as well as an enhanced representation of river discharge compared to an open-loop simulation using the former LAI minimum values. Better performances from the assimilation of disaggregated LAI are also expected on the representation of evapotranspiration.

Towards a better use of microwave satellite observations for soil moisture
ESA CCI SM is based on multiple microwave sources from space, namely passive radiometer brightness temperature (T b ) observations and active radar backscatter (σ o ) observations. As they are both indirectly related to soil moisture, retrieval methods making use of, for example, radiative transfer model (for T b , Kerr et al., 2012) or change-detection approaches (for σ o , Wagner et al., 1999) are usually required to transform T b and σ o into soil moisture values that can be assimilated in LSMs. Despite the proven record of assimilating retrieved soil moisture from point scale to regional and continental scale (e.g. Albergel et al., 2010;Draper et al., 2012;Matgen et al., 2012;De Rosnay et al., 2013;Barbu et al., 2014;Wanders et al., 2014;Ridler et al., 2014), there is an increasing tendency towards the direct assimilation of T b and σ o observations (De Lannoy et al., 2013;Han et al., 2014;Lievens et al., 2015Lievens et al., , 2016. Retrieval methods usually make use of land surface parameters and auxiliary information, such as vegetation, texture and temperature, possibly proving inconsistencies with specific model simulations (which also include these parameters but potentially from different sources). Also, if retrievals and model simulations rely on similar types of auxiliary information, their errors may be cross-correlated, potentially degrading the system performance (De Lannoy and . The direct assimilation of T b and σ o observations requires that the LSM is coupled to a radiative transfer model that serves as a forward operator for predicting σ o and/or T b . It has the advantage of allowing for consistent parameters and auxiliary inputs between the model simulations and the radiative transfer model, avoiding cross-correlated errors. The development of a forward operator for σ o from active microwave instruments is under way at Météo-France; it will allow vegetation effects to be accounted for in the signal using the vegetation information content of σ o .

Conclusions
This study provides an assessment of the LDAS-Monde implementation to increase monitoring accuracy for land surface variables over the Europe-Mediterranean area. Satellitederived surface soil moisture and leaf area index are assimilated over 2000-2012 in the CO 2 -responsive and multilayer diffusion scheme version of the ISBA land surface model coupled with the CTRIP hydrological system. Joint assimilation of leaf area index and surface soil moisture has been shown to efficiently improve the representation of aboveground biomass, gross primary production and evapotranspiration, while having a neutral to positive impact on river discharge. To our knowledge, LDAS-Monde is the only system able to sequentially assimilate vegetation products together with soil moisture observations. LDAS-Monde permits an efficient monitoring of various land surface variables and has a powerful potential in monitoring extreme events such as agricultural droughts on a global scale. The analysis of the extended Kalman filter observation operator Jacobians permitted identification of both seasonal and soil depth effects of the assimilation on ISBA. A clear added value of the assimilation has been highlighted based on agricultural statistics over France, and evapotranspiration and gross primary production observation-based estimates over the whole domain. More analysis impact could, however, be expected on variables such as river discharge. Studies focusing on a better use of the observations along with other data assimilation techniques such as the ensemble Kalman filter are currently under way. Recent studies discussed in the previous section suggest that the direct assimilation of microwave observations of T b and σ o instead of level 2 or 3 soil moisture products could leads to better results. The development of a forward operator for σ o from active microwave instruments is under way. The long-term confrontation of model and observations on continental scale prior to the assimilation has also highlighted parameterisation issues such as the minimum leaf area index values used as threshold when the vegetation is dormant. The GEOV1 leaf area index estimates permit new thresholds to be set up for the different vegetation patches used in ISBA thanks to the development of a disaggregated product resulting in new leaf area index estimates, different for each patch. The assimilation of this new product is also promising.
Code availability. LDAS-Monde is a part of the ISBA land surface model and is available as open source via the surface modelling platform called SURFEX. SURFEX can be downloaded freely at http: //www.umr-cnrm.fr/surfex/ using a CECILL-C Licence (a French equivalent to the L-GPL licence; http://www.cecill.info/licences/ Licence_CeCILL-C_V1-en.txt). It is updated at a relatively low frequency (every 3 to 6 months). If more frequent updates are needed, or if what is required is not in Open-SURFEX (DrHOOK, FA/LFI formats, GAUSSIAN grid), you are invited to follow the procedure to get a SVN account and to access real-time modifications of the code (see the instructions at the first link). The developments presented in this study stemmed on SURFEX version 8.0 and are now part of the version 8.1 (revision number 4621).