CTRIP is a physically based river routing model coupled to the ISBA land surface model (Fig. ). While ISBA represents the vertical exchanges of water and energy at the soil-atmosphere interface, CTRIP simulates river flows over an entire hydrographic network . The ISBA and CTRIP models are used, in particular, in climate models such as the CNRM-CM6 that participated in the sixth phase of the Coupled Model Intercomparison Project CMIP6 as a contribution to the IPCC Sixth Assessment Report (AR6). CTRIP is based on a regular grid with a resolution of 1/12° (i.e. around 8 km at the Equator) obtained by the upscaling of the MERIT-Hydro global hydrographic network , available at a resolution of 90 m and currently considered to be the most accurate on a global scale.

A number of hydro-geomorphological parameters, such as the lengths and slopes of river sections, are obtained from high-resolution data from MERIT-Hydro, while other parameters, such as widths, depths, and roughness, are obtained from empirical formulae . It is assumed that each grid cell contains one and only one river reach, represented as a reservoir flowing into the downstream grid cell. The Manning equation is used to calculate the flow velocity as a function of the volume of water in the section, itself updated by the inflows from the upstream reaches and the runoff from the ISBA model.

The model also benefits from a two-dimensional representation of aquifer dynamics and groundwater-river exchanges . Note that a 100-year spinup has been performed to reach a stabilized state of the aquifers. Finally, surface processes linked to vegetation (including real evapotranspiration) and snow cover (including sublimation and melting) are taken into account in the ISBA model.

It is also important to note that because the prior objective of CTRIP is to be integrated into climate models, it is based on quite simple approximations of rivers and aquifers dynamics. Consequently, unlike most hydrological models, the CTRIP model does not benefit from a parameter calibration stage. This choice ensures spatial consistency when the model is used in other regions of the world – or even globally – where few observations are available, particularly in climate projections. It follows a physically based parameterization strategy to ensure consistency across basins and scales.

The MGB model (“Modelo de Grandes Bacias” or “Large Basins Model”) is a semi-distributed hydrological model that uses physically and conceptually based equations to simulate rainfall-runoff processes and river hydraulics using a two-way coupling scheme between hydrological and hydrodynamic modules . It has been extensively applied to large river basins such as the Amazon, and has shown good ability to represent complex processes, such as floodplain dynamics .

A key feature of MGB is its explicit simulation of wetlands and floodplain processes. It includes infiltration from flooded areas into unsaturated soil columns, which is crucial in regions where dry soils are inundated by exogenous floodwaters, such as the Niger Inner Delta. The model also dynamically simulates evapotranspiration and open water evaporation based on flood-driven land cover variation, capturing the differing effects of bare soil and vegetation across flooded zones. River routing includes flow propagation, backwater effects, and hydrodynamics in large floodplains (Fig. ). A full description is provided in .

MGB is used for real-time and near real-time flood forecasting in various regions, including the Amazon basin, where it has been operational for flood forecasts and water resource management . Similar real-time forecasting applications have been implemented in the Paraná basin and other regions, showcasing the model's adaptability for hydrological forecasting at both local and large scales.

Both ISBA-CTRIP and MGB hydrological models require meteorological forcings, including precipitation, surface temperature and air humidity, wind, and radiation forcings. The choice of forcing datasets differs based on each model’s design and application.

ISBA-CTRIP is forced with ERA5 reanalysis data, which is chosen for its global coverage, temporal resolution and consistency, with its intended use for climate studies and water resource assessments . A bias-correction process of ERA5 precipitation was applied following to improve its accuracy in the hydrological context. The configuration of ISBA and CTRIP models is the one used for global scale simulations, as described in , except that the floodplain scheme was not activated. It is important to note that no observations were used to calibrate the model parameters.

In contrast, MGB is designed for basin-scale hydrology and operational forecasting. For the Niger Basin, MGB is based on the setup from , with 11 595 unit catchments and river reaches of approximately 10 km. The model is forced using GSMAP precipitation and CRU 10' long-term climatology data (monthly climate normals of wind speed, solar radiation, relative humidity, air pressure and air temperature) to calculate the model evapotranspiration. For the Congo Basin, MGB is implemented with 9220 units (20 km resolution) and forced by IMERG precipitation . These datasets were chosen for their improved spatial and temporal representation of tropical precipitation , which better aligns with the needs of short-term and flood forecasting.

The use of different precipitation forcings was deliberately retained to preserve the native configurations of the two modeling systems. CTRIP is used with bias-corrected ERA5, consistent with its global and climate-oriented configuration, while MGB is used with the precipitation products adopted in its calibrated basin-scale implementations. This choice is consistent with the scope of the study, which is to assess the ESA CCI products in realistic modelling/assimilation frameworks rather than to conduct a harmonized model-intercomparison experiment.

MGB was calibrated against in-situ discharge and evaluated using altimetry and flood extent data, while CTRIP follows an uncalibrated, physically based approach.

These differences in forcing, calibration strategy, spatial scale, and process representation are treated here as part of the native configuration of each modelling system. The dual-model setup is therefore used to examine whether the impact of long-term satellite-derived WSE and discharge assimilation is consistent across two contrasting hydrological modelling frameworks, rather than to compare absolute model performance or isolate the effect of individual model components.

Simulations are conducted over the period 2000–2020 for both basins and models, ensuring a common evaluation period for the DA experiments.

CTRIP-HyDAS (Hydrological Data Assimilation System) is a hydrological data assimilation system embedded within the CTRIP routing model. Developed since the early 2010s, it is primarily designed to assimilate observations from satellite altimetry missions. The system has been previously applied over the Amazon basin , where it was used to assimilate either nadir altimetry or SWOT-like observations. With this system, it is possible to correct model states (mainly river storage) or parameters (such as roughness).

In past studies, the assimilation of SWOT-like data was implemented in twin-experiment Observing System Simulation Experiments (OSSEs) since real SWOT observations were not available yet. Some pseudo-OSSEs have also been conducted by using a different model (namely MGB) to simulate SWOT observations, thus providing more realism in the differences between the dynamics of CTRIP and reality. More recently, CTRIP-HyDAS has been extended to operate globally at 1/12° resolution in the context of the SWOT mission. In this study, only model states are updated, using either water surface elevation (WSE) anomalies or river discharge as observation input. Assimilated data are those derived in this CCI Discharge project.

HYFAA (Hydrological Monitoring and Forecasting Framework for Assimilation Applications) is a Python-based scheduler developed to support hydrological monitoring, forecasting, and assimilation tasks using MGB model. It manages the sequantial execution and communication between three components: the MGB hydrological model, an Ensemble Kalman Filter (EnKF), and an observations database.

In its forecasting mode, the HYFAA wokflow includes three main steps:

Pre-processing, which updates local forcing and assimilation databases from external sources, converts data to the required format, and keeps track of changes in data.

Both CTRIP-HyDAS and MGB-HYFAA rely on Ensemble-based Kalman Filter (EnKF) approches to assimilate observations into their respective hydrological models. While based on the same core principles, the two systems employ slightly different variants:

CTRIP-HyDAS implements the Local Ensemble Transform Kalman Filter (LETKF), which updates the ensemble mean and perturbations separately to preserve the statistical structure of the ensemble (spread and statistical properties)

The Kalman filter is a foundational sequential data assimilation technique extensively used in hydrological sciences. At each assimilation step, it integrates observations with the forecasted state ensemble to produce an updated analysis state, effectively addressing both observational and model errors. Although originally developed for linear systems with Gaussian errors, the Kalman filter has been adapted for nonlinear models, with the Ensemble Kalman Filter (EnKF) being a notable extension .

In the EnKF framework, an ensemble of model states is propagated forward in time, and at each assimilation step, observations are used to update the ensemble members. The general update equation for each ensemble member Xi is: 1Xia=Xif+KYo+εi-HXif where Xif is the forecast state vector of the i-th ensemble member, Xia is the updated (analysis) state vector, Yo is the observation vector (here, WSE or river discharge), εi is a perturbation vector sampled from the observation error distribution, H is the observation operator mapping model states to observation space, and K is the Kalman gain matrix, computed as: 2K=PfHTHPfHT+R-1 where Pf is the forecast error covariance matrix, and R is the observation error covariance matrix, determined from the uncertainty of the measurements.

HYFAA scheduler uses the Local Ensemble Kalman Filter (LEnKF) as DA approach, which directly updates each ensemble member based on the observations and the model's error statistics.

Finally, a localization strategy has been implemented into HYFAA, as described in Sect. .

The assimilation in CTRIP-HyDAS follows the Local Ensemble Transform Kalman Filter (LETKF). The LETKF updates the ensemble mean and perturbations separately, preserving the statistical structure of the ensemble. The analysis update in LETKF is given by : 3Xa=Xf+EfVD-1VT(HEf)TRw-1Yo-HXf+m-1VD-1/2VT where Xa is the posterior state estimate, Xf is the prior state estimate (or forecast), Ef is the ensemble perturbation matrix, Yo is the observation vector, w is the weighting term used for observation localization (described in Sect. ), and VDVT is given by: 4VDVT=(m-1)I+(HEf)TR-1HEf where I is the identity matrix of dimension m×m, with m denoting the ensemble size. The matrices VD-1VT and VD-1/2VT can be calculated from the eigenvalue decomposition of VDVT. It is important to highlight that the observation error covariance matrix R is consistently assumed to be diagonal, indicating independence among observation errors.

This approach allows for efficient assimilation of observations by transforming the ensemble in a way that accounts for both model and observation uncertainties.

Both CTRIP-HyDAS and MGB-HYFAA adopt the same localization strategy to constrain the spatial influence of observations, reducing the risk of spurious long-distance correlations and improving the numerical stability of the filter. The observations localization weight w, introduced in Eq. (), is computed using the method proposed by , specifically adapted to the structure of river networks. Rather than solely relying on simple Euclidean distances or topological river-network distances, this approach defines spatial local patches based on the spatial correlation structure of hydrological variables.

For each pixel (or river reach) of the domain, a Gaussian semi-variogram is computed based on a long-term (20-year) reference simulation, used to estimate spatial correlations. The semi-variogram is fitted to time series of discharge or WSE anomalies, and its estimated range parameter determines the distance over which the variable is spatially correlated. This range is then used as a threshold to define the spatial extent of the local patch. For each observation, only the pixels within this hydrologically connected patch, both upstream and downstream, are updated during the assimilation process. This localization method is therefore more adapted to hydrology than classical methods as it accounts for the dynamic dependencies in river discharge and improves the physical consistency of the data assimilation procedure.

Figure  shows examples of the local patches for three pixels in the Congo basin, with the color bar representing the observation localization weight w.

The ensemble of model states, representing the best operating estimate, is systematically generated through the introduction of perturbed meteorological forcing. It is highlighted that only meteorological forcing is assumed to be subject to uncertainties, with the model structure and parameters assumed to be devoid of errors that could significantly impact the model outputs. Perturbing the model parameters is also a possible way to create the model ensemble, and it is particularly useful when parameters are corrected during the analysis step.

Here, we used the method developed in . The approach involves perturbing statistically significant modes obtained through the decomposition of the precipitation field into Empirical Orthogonal Functions (EOFs). Mathematically, for a variable X representing meteorological forcing (here precipitation), the decomposition can be expressed as: 5Xijk=Xmean,jk+∑ℓ=1NEOFsUℓ,jkVi,ℓ Here, Xijk represents the meteorological forcing variable at time step i, latitude j, and longitude k. Xmean,jk is the temporal mean of X at latitude j and longitude k. Uℓ,jk denotes the spatial patterns, and Vi,ℓ represents the temporal coefficients of mode ℓ obtained through the EOF analysis. The sum is taken over the retained EOFs, with NEOFs being the number of retained modes.

The leading modes, explaining 95 % of the variance, are retained, with the perturbations applied as: 6X̃ijk=Xmean,jk+∑ℓ=1NEOFsϵℓUℓ,jkVi,ℓ Here, X̃ represents the perturbed variable, and ϵℓ) is the perturbation coefficient of the respective EOF mode, following a gaussian distribution centered in 1 and with a variance of 0.4 (a value obtained by trial and error to ensure a relevant discharge ensemble spread). This methodology ensures a realistic representation of meteorological forcing uncertainties, critical for generating the ensemble of model states for subsequent assimilation experiments.

The impact of data assimilation for both models (MGB-HYFAA and CTRIP-HyDAS) across the Niger and Congo basins is first presented through a general overview of the performance changes observed across the different assimilation experiments.

To ensure clarity of the figures, only the ensemble means for each experiment are considered, given the number of assimilation experiments. It is worth noting that, for both CTRIP and MGB, the dispersion between ensemble members (standard deviation) has consistently decreased following data assimilation compared to the Open-Loop simulations.

The following initial comparison highlights the overall trends and improvements brought by the integration of observation products, which will then lead into a more detailed analysis in the following subsections. Two tables are provided in Sect. S8 in the Supplement with the detailed list of numerical performance scores shown in Figs.  and for all assimilation experiments in both basins.

In the Niger basin, when assessing the performance of MGB in assimilating discharge and WSE anomalies, Qalti shows the highest improvement in NSE (Fig. , panel 1a), with a median value of 0.83 at 50 % uncertainty, followed by WSE anomalies (NSE: 0.73–0.76). In terms of correlation, Qalti assimilation performs best, raising the median correlation to 0.94 (Panel 1b), while WSE anomalies maintain a similar value to the Open Loop (0.93). Although Qmultispec exhibits the lowest NSE (0.68–0.72) and correlation due to its higher data noisiness, it compensates by performing best in reducing bias and correcting internal variability (panels 1c,1d). Bias is best reduced by Qmultispec, improving the median from 1.2 to closer to 1 (Fig. , panel 1c). In terms of internal variability, Qmultispec also outperforms, with KGE γ nearing 1.0 at 30 %–50 % uncertainty, while Qalti and WSE anomalies range between 0.90–0.94 (Fig. , panel 1d). Therefore, while Qalti leads in improving flow magnitude and correlation, Qmultispec proves superior in bias and internal variability correction. For completeness, Fig. S11 in the Supplement shows the discharge climatologies for all stations in the Niger basin (MGB-HYFAA), illustrating the general behavior of each assimilation experiment across the full domain.

In CTRIP-HyDAS, the Niger basin exhibits limited improvement due to the model’s already poor performance in Open Loop. However, Qmultispec performs better than both Qalti and WSE anomalies in bias reduction and overall improvement of internal variability. The CTRIP model demonstrates improvements in NSE, where Qmultispec assimilation leads to a 15 % increase in performance, whereas WSE and Qalti show more modest gains in both bias and correlation (panels Fig. ).

For the Congo basin, assimilation experiments reveal a more nuanced picture. The analysis is limited to only two of the four main tributaries and Kinshasa at the lower Congo, making the scope of conclusions more constrained. While MGB displays noticeable improvements with Qmultispec in the Niger basin, the product does not perform as well in the Congo basin, largely due to lower data quality and temporal density (Fig. S2 in the Supplement). Here, Qalti delivers the best performance, notably improving NSE from 0 to 0.1 at Kinshasa. Correlation scores improve by about 5 % at Kinshasa and other stations (Panels 1a, 1b, Fig. ). However, bias and internal variability are only slightly corrected (less than 5 % overall improvement), as seen in panels 1c and 1d of Fig. .

Meanwhile, in CTRIP, the assimilation of WSE anomalies leads to better performance over the Congo basin compared to discharge assimilation. This is due to a more coherent alignment between the model’s rating curve and observed WSE data (as shown in the following Sect. ), along with the lower uncertainty attributed to WSE compared to discharge. Panels 2b and 2d of Fig.  show improvements in correlation (+13 %) and variability (-7 %), particularly noted at Kinshasa, where Qmultispec underperforms. In contrast, Qmultispec data quality performance in both MGB and CTRIP for the Congo basin, particularly at Bangui and Kinshasa, where it fails to accurately represent the hydrological cycle.

When comparing the influence of the observations characteristics (spatial density, temporal frequency, and data accuracy) with uncertainty values, the distribution of performance indices across experiments shows that the characteristics of observed data play a more decisive role in both models and basins. Data assimilation is primarily driven by the spatial density of virtual stations, followed by temporal resolution and data accuracy, with uncertainty having a lesser impact. However, uncertainties become important in the presence of outliers, as higher uncertainty values reduce the likelihood of outlier assimilation, minimizing hashed effects in the analysis. This is evident in the assimilation of Qmultispec products in the Congo basin, where poor data quality degrades model performance. Nevertheless, higher uncertainty (50 %) limits degradation compared to 30 % uncertainty (Fig. ).

From these results, two key findings are highlighted to explore more specifically the contributions of the different CCI discharge products. In one hand, the comparison between discharge assimilation and WSE anomaly assimilation in large-scale hydrological models is examined. In the other hand, the added value of Qmultispec data is analyzed, particularly in the Niger Basin, where it outperforms Qalti in correcting bias and internal variability.

The comparative performance of dH and Qalti assimilation is influenced by several factors, including product uncertainty, model open-loop baseline simulation, and how well observations integrate with model dynamics. Among these, the mismatch in rating curves between models and observations emerge as a key recurring limitation for dH assimilation and is highlighted at relevant points throughout this section.

In the Niger Basin, MGB shows a clear advantage for discharge assimilation (Qalti) over water surface elevation anomalies (dH). Qalti improves Nash-Sutcliffe Efficiency (NSE) to 0.83 (+14 %) and raises correlation to 0.94 (+18.5 %), ensuring smoother flow predictions across stations.

As shown in Fig. , in Ansongo and Malanville, discharge assimilation produces smoother results, while WSE assimilation introduces noise, with visible jumps in the simulated discharge after each update. These hashed effects arise when corrections are applied intermittently, and the model gradually returns to its prior state between updates. They are especially pronounced when observations are temporally sparse and when there are large discrepancies between observed and simulated values.

The hashed effect is further amplified when assimilating WSE at stations with a strong mismatch between observed and simulated rating curves. At Ansongo, for example, this mismatch (Fig. ) leads to conversion errors when using WSE anomalies to update discharge, amplifying the hashed effect. In contrast, these are much less pronounced at stations like Malanville (Fig. ), Ibi, and Lokoja, where the model and observation rating curves from the model and observation are more consistent (Fig. , panel b; Fig. S10 in the Supplement). These stations also benefit from overlapping local patches that include observations from nearby stations (e.g., Malanville includes Niamey; Lokoja includes Malanville and Ibi), which helps sustain corrections between assimilation steps. However, the overall effectiveness of the local patch remains constrained by the limited spatial and temporal coverage of the CCI products. To provide further context, Figs. S9 and S10 in the Supplement illustrate respectively the local patches for the six stations in the Niger basin, and rating curve comparisons between MGB (Open-Loop) and CCI observations (dH, Qalti) at Ibi and Lokoja. Finally, because Qalti is a directly modeled variable, it integrates more consistently into the hydrological model equations. This contributes to the smoother corrections observed across all stations during its assimilation, as illustrated for example at Ansongo and Malanville (panels (2b) and (2d), Fig. ), and similarly seen the rest of the stations.

In addition to these structural mismatches, such as those seen in Ansongo, the behavior of the assimilation filter is also shaped by how observation uncertainties are handled in the EnKF. At Ansongo, the MGB model results for two experiments (dH 0.4 m and Qalti 30 %) from 2010–2020 reveal additional differences in assimilation, as shown in Fig. . Panels (1b) and (2b) represent the time series of increments, defined as the difference between the model analysis and the background (i.e., the forecast prior to assimilation). These increments quantify how much the model state was updated by the assimilation process. Panels (1c) and (2c) show the residuals, defined as the difference between the analysis and the assimilated observations: a low residual indicates the model is closely following the observations, while a large residual suggests a partial rejection of the observation. Finally, the innovation (not explicitly shown here) would be the difference between the observation and the background (forecast), and shows the initial discrepancy before assimilation.

With dH assimilation, the model almost entirely follows the observations, as indicated by the near-zero residuals. The same pattern is observed for CTRIP. Due to their lower uncertainty compared to the other products, dH observations, even at 0.4 m uncertainty value, are still assimilated even when they diverge from the model's theoretical WSE values. This introduces erroneous information into the model, which can disrupt the simulated discharge and amplify the hashed effect after assimilation.

Qalti, in contrast, despite its 30 % uncertainty, results in larger residuals, especially during high flow periods. This is mainly because the error given to the discharge products is proportional, so the higher the discharge value, the higher the absolute uncertainty given to that value and the lower is the weight given to observation within the EnKF scheme compared to model background. These residuals help maintain more reliable results by preventing the full assimilation of outliers (Fig. ). This selective assimilation behavior of Qalti, driven by proportional uncertainty, is particularly useful for reducing the influence of extreme outliers during flood peaks. However, during low-flow periods, Qalti's uncertainty becomes very small in absolute terms, potentially leading the filter to overfit isolated or erroneous low-flow observations. This limitation should be considered when evaluating Qalti assimilation performance across all flow regimes. dH is less sensitive to this effect due to its fixed absolute uncertainty.

Another key advantage of Qalti lies in its temporal continuity, since discharge is a mass-conserving quantity that evolves continuously over time steps, supporting more physically consistent updates throughout the river network. In contrast, dH values are tied to local rating curves and may produce temporally inconsistent discharge increments between time steps and across stations. A 0.4 m WSE anomaly may lead to a discharge correction that varies widely across the river network, depending on the local shape of the rating curve. In reaches with steep slopes or narrow channels, this may correspond to small discharge changes. In contrast, in flat floodplain areas or during low-flow periods, this can induce disproportionately large discharge corrections, potentially destabilizing the model.

In the Congo Basin, the mismatch between the MGB model’s rating curve and the observed altimetry data is even more pronounced, especially at Kinshasa. At this location, Qalti outperforms dH, as dH assimilation fails to capture the large observed WSE variations due to the model's flat rating curve (Fig. ). At Kinshasa, the MGB model has a flatter slope, meaning that for similar discharges, the model produces much smaller changes in water surface elevation. In contrast, the observed rating curve from altimetry (dH and Qalti) shows a much steeper slope, indicating a larger variation in water surface elevation for the same discharge fluctuations. Altimetry observations show dH variations reaching up to ±7 m, while the MGB model’s maintains nearly constant WSE values. This mismatch between modeled and observed rating curves, also noted in the Niger Basin (Ansongo, Fig. ), illustrates a broader structural limitation of dH assimilation in MGB: when rating curves slopes differ significantly, converting WSE anomalies to discharge can introduce significant errors.

This mismatch leads to EnKF filter divergence, resulting in outliers and highly negative NSE values, which makes dH assimilation unreliable at stations such as Kinshasa and Ouesso (Fig. ). Ouesso, correlated with Kinshasa through the localization matrix, is also impacted by the discrepancies observed at Kinshasa. These inconsistencies propagate backward to Ouesso, resulting in unstable discharge estimates with extreme fluctuations, including both excessively high and negative values. These issues lead to poor performance scores at Ouesso as well. It is worth noting that, in this version of MGB-HYFAA, the EnKF does not impose numerical constraints on discharge, which allows for negative values when unrealistic corrections deviate from physical plausibility, as seen in Fig. .

While MGB benefits more from Qalti due to its compatibility with discharge as the primary modeled variable, CTRIP shows a different behavior. The assimilation of dH consistently outperforms Qalti across the stations in both basins. This difference is most notable in the Niger Basin, where the Open-Loop (OL) simulation performs poorly. In such cases, the EnKF favors the observation with the lowest uncertainty, typically dH, even when Qalti would be more consistent with the model's output. Since Qalti is indirectly derived from altimetry and involves transformation through rating curves, it carries higher uncertainty and is thus assigned lower weight in the analysis. This may lead to over-reliance on dH, even if the structural limitations (e.g., flat rating curves) make it less appropriate.

dH assimilation is most effective in CTRIP when large model biases are present and when model rating curves align well with observations. In fact, CTRIP’s rating curves closely match altimetry-based ones at most stations, including Niamey, as shown in Panel (c), Fig. . Panel (a) shows that dH assimilation shows consistently higher KGE scores. Given equal temporal resolution, the lower uncertainty of dH (0.2–0.4 m) gives the observation more relative weight in the analysis, compared to Qalti (30 %–50 %).

However, this comes at the cost of physical consistency and temporal continuity. Q is a mass-conserving variable that evolves continuously over time steps, supporting more physically consistent updates throughout the river network and therefore offering more stable integration with hydrological model dynamics. WSE anomalies, by contrast, do not conserve volume and require geometric assumptions through rating curves, which can reduce consistency between time steps.

In CTRIP, where structural errors, like missing evaporation downstream the Inner Delta in the Niger, lead to biased flow, dH assimilation leads to better improvements, but only when the model's rating curve matches observations (Fig. ).

In contrast, MGB includes explicit evaporation and provides a better OL simulation. In this case, Qalti proves more effective: it updates discharge directly and avoids transformation errors, and better reflects seasonal flow signals. Even at stations with good rating curve alignment (e.g., Malanville), Qalti outperforms dH assimilation (Fig. ). This is supported by comparisons with in-situ data (Sect. S1 in the Supplement), which show CCI discharge products capturing seasonal dynamics reliably.

These results indicate that discharge assimilation tends to be more effective when the model provides a reliable physical representation of the river system and the open-loop is not structurally too poor.

The preferred assimilation variable thus depends on a combination of the model's open-loop performance, the uncertainty of the observation, and rating curve alignment. In CTRIP, structural limitations and accurate alignment with observed rating curves favor WSE assimilation. In contrast, Qalti assimilation in MGB is supported by stronger internal discharge simulation and more realistic physical representation. The issues with rating curve mismatches in MGB (Kinshasa, Fig. ) and the close performance results between dH and Qalti in CTRIP (Fig. ) highlight the sensitivity of each variable’s performance to model context.

The choice between dH and Qalti should consider not only uncertainty but also how well each variable aligns with the model’s structure and dynamics. For large-scale hydrological models, whose primary objective is to accurately simulate discharge rather than water surface elevation (WSE), it is more robust to assimilate a variable directly related to the flows they simulate. Adjusting rating curves across model cells to accommodate dH is more complex and introduces errors, especially when the model’s simplified river geometry does not match real-world conditions. Discharge assimilation avoids these conversions and ensures consistency across both model cells and time steps, making it the more reliable choice overall. Thus, when the model is not too structurally biased, Qalti assimilation tends to outperform dH and is better suited for large-scale hydrological applications.

In the Niger Basin, despite its lower performance in other metrics, Qmultispec outperforms in bias correction by significantly reducing flow overestimation in both models. This improvement is particularly notable in the upper Niger and Inner Delta within MGB, where the Open-Loop (OL) simulation consistently overestimates discharge. Koulikoro serves as a key example, where the bias is reduced by 17 %–20 % (depending on the uncertainty levels), bringing the bias from 1.2 in the Open-Loop down to approximately 1. In fact, Qmultispec’s higher temporal density enables better alignment with the observed rising and falling limbs of the hydrograph, providing more consistent results over the annual cycle.

Figure  shows time series at Koulikoro (2010–2020, MGB) for two experiments at 30 % uncertainty: Qalti (1a) and and Qmultispec (2a). Panels (1b, 2b) show the time series of increments (analysis minus background). With frequent updates, bias corrections occur steadily throughout the hydrological year, correcting both high- and low-flow biases.

Panel b of Fig.  confirms that Qmultispec produces smaller mean increments during the high-flow period (June–November) compared to dH or Qalti assimilation. At Koulikoro, Qalti actually overestimates discharge by 35 % compared to in-situ data, while the MGB Open-Loop overestimates Open-Loop overestimates by 25 %, and Qmultispec by 20 %. Although the Qmultispec data is noisier, particularly during the recession phase, its frequent updates prevent the model from reverting to its biased pre-assimilation state.

In the CTRIP model, the trend is similar: Qmultispec provides the greatest reduction in bias, followed by dH and Qalti. This difference comes from the lower uncertainty levels of the WSE anomalies (0.2–0.4 m) compared to Qalti (30 %–50 %). In the Niger Basin, CTRIP’s Open-Loop simulation overestimates discharge significantly, so all satellite products help reduce this positive bias. The frequent assimilation from Qmultispec helps constraining the model and reduces the tendency to revert to biased states (Open-Loop). This is particularly visible in the monthly climatologies, as shown at Niamey in Fig. .

Beyond bias correction, Qmultispec also improves internal variability, particularly in MGB. With dense temporaly data coverage, it enables dynamic corrections that reduce simulation divergence caused by model internal errors or the omission of complex physical processes. This is evident in both the MGB and CTRIP models, as seen in Fig. . It is worth noting that within the Niger basin, the internal variability is less effectively corrected by any of the products in CTRIP, largely due to the model’s poor Open-Loop performance, which limits its ability to capture short-term fluctuations.

Qmultispec’s improved capture of short-term fluctuations makes it particularly valuable for climate studies focusing on flow variability and its evolution over time. Capturing this internal short-term variability is essential for modeling extreme events such as floods and droughts, which makes this product doubly relevant for MGB, as it is designed for forecasting applications.

Overall, in both models, Qmultispec outperforms Qalti and dH in reducing bias and correcting internal variability. Despite some data noisiness, Qmultispec delivers more Qmultispec delivers more consistent and reliable performance compared to Qalti and WSE anomalies. This makes it particularly valuable for long-term studies on the evolution of discharge and water resources, such as those related to climate change, as well as for forecasting extreme events.