Given the required revisit frequency and the Ucayali River's width in the plains (500–1000 m), moderate-resolution satellite imagery from MODIS (MODerate Resolution Imaging Spectroradiometer, 250 × 250 m, 1999–present, 1–2 d) and VIIRS (Visible Infrared Imaging Radiometer Suite, 375 × 375 m, 2012–present, 0.5 d) was used to generate time series of surface water reflectance (Fig. 1c). Reflectance values in the red and near-infrared (NIR) bands were extracted pixel by pixel from satellite images using the free software GetMODIS and MOD3R, developed by the CZO HyBAm, that was tested and validated in various previous studies (e.g. Espinoza-Villar et al., 2017; Vauchel et al., 2017). Water masks were applied to the Ucayali River's main course near virtual stations. To ensure 50–100 pixels per mask, large river stems were covered, with masks redrawn every 2–3 years due to river mobility. Collected scenes comprise images spanning 8 d periods, selecting pixels with the lowest cloud cover and smallest satellite-viewing nadir angle.

Two radiometric campaigns were conducted in the Ucayali Basin: the first in November 2011 at Requena (Espinoza-Villar et al., 2012) and the second in February 2017 at Lagarto, Puerto Inca, and Pucallpa, spanning three weeks (Santini, 2020). A total of 42 surface water samples were collected to determine total, fine, and sand concentrations. Simultaneously, hyperspectral field radiometers (TriOS) were deployed following the experimental setup of Mobley (1999), as adapted by Martinez et al. (2015) for the Amazon Basin. High-frequency (1 Hz) hyperspectral measurements of surface water reflectance were obtained at sampling locations. Relying on this dataset, a unique model for all the Ucayali Basin was fitted between fine sediment concentration at the water surface and the ratio of radiometer reflectance in the NIR (841–876 nm, according to the satellite sensor bands) and red bands (620–670 nm) (see Sect. 5.4). This single relationship is considered applicable along the mainstem of the Ucayali River in the lowland plain because surface reflectance is largely controlled by fine silts of Andean origin, whose grain-size characteristics and optical properties are relatively homogeneous along the river continuum (Martinez et al., 2015; Santini, 2020). In addition, the use of spectral band ratios (NIR / RED), commonly applied to reduce the influence of these potential variations in optical conditions (Doxaran et al., 2002; Martinez et al., 2015; Pinet, 2017), supports the applicability of this relationship at the basin scale considered in this study.

Due to the considerable depth of Amazonian rivers and the vertical sediment concentration gradient near the surface, the ratio αf, relating the channel mean concentration to the surface index concentration retrieved by satellite, ranges from 1 to 1.8 according to the CZO HyBAm database (1.1 to 1.2 in the Ucayali). These values are derived from field measurements of vertical suspended sediment concentration profiles collected at several stations of the HyBAm observatory across the Amazon Basin, including the Ucayali River (e.g. Santini et al., 2019; Santini, 2020). To estimate αf along the river network, the models proposed by Santini et al. (2019) were applied. These models use Rouse-type formulations constrained by observed concentration profiles to describe the vertical distribution of suspended sediments as a function of hydraulic conditions. They were parameterized using hydraulic variables simulated by SWAT-Amazon.

The mean water level h (m) is derived from the water volume V (m³) stored in the reach i at time step t (beginning of the simulation day). As long as h≤hf, where hf (m) is the floodplain activation threshold, h is computed as: 1hit=VitΔxiBi, where B (m) is the main channel width and Δx (m) is the reach length. When h≥hf, the floodplain is activated, distributing V between the main channel and floodplain. If the floodplain cross-section is rectangular: 2hit=Vit-ΔxiBihf,iΔxiBi+Wfp,i+hf,i, where Wfp=kfpB (m) is the floodplain width and kfp (–) a coefficient to be calibrated (the SWAT-Amazon parameters are given in Table 1). For a triangular floodplain cross-section, h depends on θfp (rad), the riverward slope angle (see S4 in the Supplement). Finally, state variables such as wetted area Ah (m²), wetted perimeter Ph (m), and hydraulic radius Rh (m) are all derived from h.

In large Amazonian rivers, flow variations over time and space are minimal, leading to a subcritical hydraulic regime, that can be modelled using the 1D Barré de Saint-Venant equations (Moussa and Bocquillon, 2009). Given the (very) gradual flow variations (Trigg et al., 2009), the convective and local acceleration terms are negligible, making the diffusive flood wave approximation suitable. When water-surface slope effects are also negligible, the pressure gradient is eliminated, allowing the use of the kinematic wave equation. SWAT-Amazon enables reach-specific selection between kinematic wave (Sf=Sb), suitable for steep Andean reaches, and diffusive wave approximation (Sf=Sb+Sw), preferred for low-slope floodplain reaches where backwater effects may occur (e.g. Yamazaki et al., 2011), with Sf (m m⁻¹) the energy gradient (or friction slope), Sb (m m⁻¹) the bed slope and Sw (m m⁻¹) the water-surface slope. When the diffusive wave model is used, Sw is assessed as follows: 3Sw,it=hi+1t-hit12Δxi+Δxi+1. The reach-averaged velocity u (m s⁻¹) and discharge Q (m³ s⁻¹) are then calculated using the Gauckler-Manning-Strickler (GMS) friction equation (Hager, 2005). 4Qit=Ah,ituit=Ah,itSf,itnc,itRh,it2/3. In the present study, the diffusive wave formulation was applied throughout the river network in order to account for backwater effects in the low-slope reaches of the Ucayali River.

At the end of the calculation time step (t+Δt), V in reach i is updated as: 5Vit+Δt=Vit+Qi-1t-QitΔt+VRit-VEit-VTit, where VE (m³) is the volume lost to evaporation, VT (m³) is the volume infiltrated into the unsaturated water table and VR (m³) is the runoff (surface, subsurface, and baseflow) reaching the river. The computation of VE, VT and VR follows the standard SWAT model. The updated volume Vit+Δt is then used to determine hit+Δt for the next simulation step.

Although SWAT simulations are reported at a daily time step, river routing is internally solved at a sub-daily time step dynamically determined by the Courant–Friedrichs–Lewy (CFL) stability condition. The routing module automatically adjusts the internal time step to satisfy the CFL criterion and ensure numerical stability (Bates et al., 2010). Daily outputs therefore correspond to the model state at the end of each simulation day.

At the beginning of the simulation day, the suspended sand concentration Cs (t m⁻³) in the reach i is: 6Cs,it=Vs,itQituitΔxi, where Vs (t) is the sand volume stored in the reach, in the main channel only. The daily suspended sand load Qs (t d⁻¹), taking Δt= 86 400 s, is then: 7Qs,it=QitCs,itΔt,

Selecting appropriate transport capacity equations for deep, low-gradient rivers is crucial, as most were derived from laboratory studies under opposite conditions (steep slopes, shallow water, uniform flow). The physically based Camenen and Larson (2005, 2008) models for non-cohesive sands were chosen for their calibration with extensive global datasets and proven applicability in large tropical rivers (Camenen et al., 2014). In these models, the transport capacity Qs∗ (t d⁻¹) for suspended sands is evaluated as a function of the Rouse number Ps (–), which defines the concentration profile exponential shape, and a near-bed reference concentration Cb∗ (m³ m⁻³), which determines its magnitude: 8Qs∗it=Cb∗it16Psit1-exp⁡-6PsitQitsΔt, where s is the relative sand density. The Rouse number summarizes the equilibrium between grain settling velocity ws (m s⁻¹) and turbulence-induced lift, related to the shear velocity u∗ (m s⁻¹), weighted by the sediment-to-eddy diffusivity ratio βs (–): 9Psit=ws,iβs,itκu∗it, where κ is the Von Kármán constant. The Soulsby (1997) law is used for estimating the sand grain settling velocity, involving the grain size ds (m) of the suspended sands. The shear velocity is calculated using the depth-slope product: 10u∗it=ghitSbit+Swit, where g (m s⁻¹) is the gravitational acceleration. The diffusivity ratio is either assigned a fixed value for each reach or computed dynamically using the Santini et al. (2019) model: 11βs,it=3.1exp⁡-0.19×10-3u∗itws,ihitds,i0.6+0.16, The bottom reference concentration Cb∗ is given by Camenen and Larson (2005): 12Cbi∗t=0.0015θ′itexp⁡0.2db,i+4.5θcr,iθ′it, where θ′ is the dimensionless grain-related bed shear stress and θcr (–) the critical Shields parameter for the inception of transport (Camenen et al., 2014), which can be estimated from the Yalin-Shields curve as a function of the riverbed dimensionless mean diameter db∗: 13θcr,i=0.25db∗,i+0.0551-exp⁡-0.02db∗,i. The parameter θ′ is calculated as following: 14θ′it=12f′ituit2gs-1db,i, where f′ (–) is the Darcy-Weisbach skin roughness factor, derived from the logarithmic velocity law: 15f′it=2κln⁡30ks,i′hit-12, with κ is the Von Kármán constant and the height ks′ (m) is the hydraulic skin roughness of Nikuradse, which can be expressed as a function of db and a coefficient σ=2.5 (Engelund and Hansen, 1967; Bartholdy et al., 2010): 16ks,i′=σidb,i.

The difference ΔQs∗ (t d⁻¹) between the transport capacity Qs∗ (t d⁻¹) and the sand load Qs (t d⁻¹) is then calculated at time t: 17ΔQsi∗t=Qsi∗t-Qsit. If ΔQs∗>0, there is excess transport capacity, allowing for riverbed erosion. The eroded mass Ebed (t d⁻¹) is defined as: 18Ebedit=KbediΔQsi∗t, where Kbed (–) is a coefficient (0≤Kbed≤1) representing the susceptibility of the channel bed to erosion when the simulated sand transport capacity exceeds the available sand load. This parameter governs the potential entrainment of bed material into suspension and thus the possible contribution of riverbed erosion to the simulated suspended sand flux.

The sand flux is then updated: 19Qsit=Qsit+Ebedit. Conversely, if ΔQs∗≤0, the sand load exceeds transport capacity, and the sand load is set to the transport capacity: 20Qsit=Qsi∗t.

Drawing on the mass balance proposed by Dunne et al. (1998), an erosion term, Ebk (t d⁻¹), is introduced to account for both floodplain channel inputs and bank erosion. These processes primarily occur at point bars on the inner bends of meanders, where floodplain inflows, with lower sediment concentrations than the river's transport capacity, enhance erosion and resuspension. Riverbed erosion, Ebed (Eq. 18) and two deposition terms (Dovbk, Dlat), all expressed in (t d⁻¹), are also considered. Riverbed erosion is therefore explicitly represented in the model formulation so that its potential contribution to the suspended sand flux can be evaluated during calibration. Sand deposition on bars in low-velocity zones of the main channel is already accounted for in the sand load adjustment (Eq. 20). Deposition in floodplain channels and levee depressions when active (i.e. when h>hf) is neglected for sand particles, as the high flow resistance caused by vegetation in these areas is expected to result in complete sedimentation at their inlets. This process is therefore implicitly included in the overbank deposition term, Dovbk.

The term Ebk is activated only when the daily water volume ΔVfp (m³ d⁻¹) exchanged between the main channel and the floodplain is negative, meaning floodplain waters contribute to the main channel. Thus, Ebk was defined as function of ΔVfp and Cbk (t m⁻³), the concentration of these banks and bars inputs, considered as a constant to be calibrated: 21Ebkit=-ΔVfp,iCbkit. Thus, Ebk is neglected when ΔVfp≥0 because the volume of water that could flow back from the floodplain during the rising stages is low compared to the water discharge in the main channel, contrary to the flood recession phase. In addition, the term Ebed can compensate for this if necessary, when the transport capacity is in excess.

The daily sand mass Dovbk is defined as a function of ΔVfp: 22Dovbkit=ΔVfp,iCsit(zsurfit), where Cszsurf (t m⁻³) is the sand concentration in the upper flow layer, estimated at zsurf≅h-hf/2 (m). Cszsurf is derived from the mean concentration Cs in the reach: 23Csit(zsurfit)=Csitαs. The ratio αs=Cszsurf/Cs is estimated with the Santini et al. (2019) model: 24αsit=16Ps,cfitexp⁡6Psitzsurfithit1-exp⁡-6Psit.

At the end of the calculation time step, the sand volume Vs stored in the main channel of the reach i is updated as: 25Vs,it+Δt=Vs,it+Qs,i-1t-Qs,itΔt-Dovbk,it+Ebk,it, As for water routing, sand routing is internally solved at a sub-daily time step to satisfy the CFL stability condition. The routed sediment inflow provided by the SWAT-Amazon reach network represents the sediment supply entering each reach and is incorporated into the reach-scale sand mass balance described in Eq. (25).

When the floodplain becomes active, differences in depth and roughness between the main channel and floodplain develop a shear interface between the two flow zones, associated with Kelvin-Helmholtz instabilities, transferring horizontal momentum from the main channel to the floodplain (e.g. Sellin, 1964; Nicollet and Uan, 1979; Ervine and Baird, 1982; Knight, 1989; Knight and Shiono, 1996; Smart, 1992; Loveless et al., 2000; Yen, 2002; Uijttewaal, 2014; Atabay and Knight, 2018; Proust and Nikora, 2020). Sediment-laden water flowing through floodplain channels (Lewin et al., 2017) also transfer large amounts of momentum to the plain and reduces the kinetic energy of the main flow, as does the attenuation of the water surface slope during flooding, which tends toward the valley slope. Moreover, the waters that travel for a short time through the floodplain before returning to the main channel also contribute to reduce the flow velocity. These combined effects significantly reduce flow velocity and, more drastically, transport capacity in the main channel. They change the spatial distribution of velocities and shear stress in the main channel cross-section, especially near the banks and bars, where sediment stocks can be available. To account for this, a flow resistance correction factor ζn was defined as: 26ζn=ucfuc=ncncf, where the subscripts “c” and “cf” denote in-bank flow configuration (without floodplain) and flow with an active floodplain, respectively, at the same water level h>hf. To evaluate ζn, the Nicollet and Uan (1979) or Smart (1992) equations can be used. However, both formulations only consider the shear layer interface between the main channel and floodplain. Furthermore, the Smart equation is not suitable for large rivers, and the Nicollet et Uan equation requires an estimate of the floodplain's Manning coefficient. Although the latter was implemented in the new water routing module, a simpler approach was preferred. Therefore, a relationship between ζn and the relative height Y=h-hf/h (–), from which the water exchanges between the main channel and the floodplain begin to affect the flow velocity, was defined: 27ζnit=11+YitCnfp,i, with Cnfp (–) a coefficient to be calibrated, superior to zero if the flood impacts the flow resistance. Thus, to account for the floodplain drag when h>hf, the Manning coefficient is reevaluated as follows: 28ncfit=niζnit. In the calibrated model, Cnfp values along the mainstem range between 0.3 and 1. Under typical flood conditions, Y varies between 0 and about 0.3, resulting in an increase of the effective Manning coefficient of approximately 0 %–6 % (see S5 in the Supplement). Larger corrections, up to about 20 %, may occur only during extreme simulated floods when Y approaches 0.5. Such variations are consistent with resistance changes reported in studies of compound channel hydraulics and floodplain–channel momentum exchange (e.g., Nicollet and Uan, 1979; Smart, 1992; Knight and Shiono, 1996; Bousmar and Zech, 1999). The coefficient Cnfp was calibrated during the model calibration phase to reproduce observed stage–discharge relationships at the mainstem gauging stations.

Following the same reasoning as for velocities the ratio of the dimensionless grain-related bed shear stresses θc′ and θcf′ should also be a function of ζn. Indeed, according to Eq. (14): 29θcf′θc′=fcf′ucffc′uc2=ζn2. Here, fcf′=fc′ is assumed, as f′ is a grain-related friction factor, not a flow resistance factor (Yen, 2002): the floodplain drag is already accounted for in ucf, through ncf. The shear velocity term used for calculating the Rouse number (Eq. 10) should be also affected by the floodplain drag: 30u∗cf=ζnu∗c. However, when shifting from a 1D to a 2D framework, the transverse profiles of θ′, and consequently of Cb∗ and Qscf∗, are likely to be more strongly affected than the lateral profile of the depth-averaged velocity (see S6 in the Supplement), in particular near the banks. To account for the complex 2D effects on sediment transport capacity, effects not considered in the initial computation of the transport capacity Qscf0∗ which was initially calculated using the corrections for θ′ and u∗ corrections in Eqs. (29) and (30), the following formulation is applied when the floodplain is active (i.e. when h>hf): 31Qs∗cfit=ζnηitQs∗cf0it, with η an exponent to calibrate which accounts for these complex 2D effects.

In the large Amazonian rivers, a decrease in bed roughness influence with increasing water levels has been observed (see example in S7 in the Supplement). To model this in the SWAT-Amazon version, the Manning coefficient is modified using the factor ζnch, defined as: 32ζnchit=1+Cnchihchi-hithit, where Cnch (–) is a coefficient and hch (m) the water height below which the additional bed roughness correction activates. Calibrated values of hch range between 35 % and 75 % of the maximum simulated water level across the mainstem reaches, and Cnch between 0.12 and 0.4, resulting in an increase of the effective Manning coefficient of approximately 10 %–30 % at low stage, decreasing progressively as stage rises above hch (Fig. S4). Both parameters were calibrated to reproduce the observed stage–discharge and stage–velocity relationships at the mainstem gauging stations.

The analysis focused on the parameters set (n, B, hf, Cnfp, kfp), with n accounting for all the resistance term Sb/n. Results show a greater sensitivity of the model to (hf, Cnfp, kfp), which control flood wave attenuation by the floodplain (Fig. 9a). The interannual time-series analysis (Fig. 9c) reveals greater sensitivity during flood recession than rising waters. The recession is particularly challenging to calibrate due to rapid water discharge drops, where even slight shifts in the timing of floodplain recession cause large discrepancies between simulated and observed flows throughout the entire recession limb. Therefore, the parameter hf, controlling recession onset, must be carefully assessed. The strong oscillation of Sobol indices for n during low waters (Fig. 9c) reflects the impact of h (through B) on bed roughness and flow resistance when h<hch (cf. Eq. 32). Small variations of h induce large changes in n during low waters, but with minimal impact on discharge, as shown by the interannual discharge plot in Fig. 9b. This effect diminishes quickly as water levels rise.

The sensitivity analysis, using the parameter set (βs, Cbk, db, ds, η, kbed), shows that db is the most critical calibration parameter (Fig. 9d, f), as previously highlighted by Fagundes et al. (2023). In the present framework, db and ds are assumed to remain constant over the seasonal cycle. Calibrated db values for sub-basins 19, 5, 14, 23, 22, and 21 are 252, 240, 240, 242, 235, and 220 µm, respectively, matching observed mean diameters at Lagarto (260 µm), Pucallpa (243 µm), and Requena (228 µm). Calibrated ds values are approximately 80 µm for all sub-basins, except Lagarto (98 µm), are consistent with PSD observations. Since the calibration of the flood recession limb directly influences sand resuspension (Eq. 22), hf emerges also as a key parameter in sand routing. The influence of kbed is minimal, as the sand load input in sub-basin 21 is sufficient, eliminating the need for riverbed erosion to compensate for sediment supply deficits.

GLUE-based dotty plots were produced by sampling the Sobol-influential parameters over wide ranges (S11 in the Supplement). The water routing parameters (hf, B, kfp, Cnfp, n) all show identifiable optima to varying degrees depending on the reach, consistent with the Sobol analysis. Critically, db exhibits a physical optimum near 220–250 µm matching observed grain sizes, and spurious coarser modes, precluding sand suspension. This bimodality highlights a well-known limitation of purely statistical calibration: without physical constraints, optimization can converge toward parameter values that are mathematically acceptable but physically meaningless. Overall, since db is well constrained and water routing parameters are collectively identifiable, the hydro-sedimentary budgets remain robustly constrained despite residual equifinality across all behavioural simulations (S12 in the Supplement). These findings justify the use of expert measurement-informed priors for uncertainty propagation.

According to the previous analyses, the calibration strategy for stations with robust, long-term hydro-sedimentary monitoring is:

a.

Start by calibrating Q in each reach, for hf only, using the SWAT's default hydrologic parameters.

It is important to emphasize that the optimal calibration for water discharge may not align with the best calibration for water level, velocity, and sand load time series. A compromise must be made.

The sensitivity of the model to n and B emphasizes the importance of h (Eqs. 1, 2, 4) in the routing module, particularly for floodplain inundation and dewatering triggers (Sect. 4.1), as already discussed above. For ungauged reaches (5, 23, 22, see Fig. 2a) in the plain, n and B can be calibrated using relative water levels (hydraulic radius) derived from spatial altimetry and shifted by a constant offset to match simulations (Fig. 10). The Manning coefficient is supposed already known to a reasonable degree, as coherent values of 1/45 ± 0.0005 s m^−1∕3 have been found across all the other plain's sub-basins. Given the large sub-basins and significant channel lengths, bed slope errors are assumed to be smoothed, and therefore Sb also well assessed. Thus, velocities can be reliably determined if B is calibrated through h calibration. The calibration for reaches (5, 23, 22) was also guided by looking for the best simulation at Pucallpa and Requena.

Finally, implementing an integrated approach combining modelling and remote sensing for monitoring hydro-sedimentary fluxes would require revisiting hydrological network operations. Specifically, a few gauging campaigns (around four, strategically timed along the annual hydrograph) could be sufficient for model calibration. Additionally, synchronizing gauging with satellite altimetry would enable the construction and simulation of robust rating curves.

Based on the long-term monitoring experience of the CZO HyBAm network and the results obtained in this study, the framework can be applied to other Amazonian basins provided that a minimal observational dataset is available at a limited number of super stations and complementary virtual stations. For the last ones, this includes: (i) approximately four to eight discharge gauging campaigns covering the main stages of the annual hydrograph (rising limb, high water, falling limb, and low water), (ii) a comparable number of suspended sediment gaugings distributed along the hydrograph and carried out using standardized and hydraulic-based protocols (Santini et al., 2019), (iii) one to two detailed PSD surveys in relatively homogeneous upstream reaches, increasing to four to five in more heterogeneous downstream settings (e.g. below major confluences), and (iv) two to three radiometric calibration campaigns spanning contrasted hydrological conditions.

This study adopts a hydraulic representation of the river, modelling the floodplain as a reservoir. Rapid exchanges through secondary channels are neglected, and their influence is approximated on discharge propagation through a minor adjustment of the Manning's coefficient. Guilhen et al. (2022) used a previous version of SWAT-Amazon with a slightly modified floodplain configuration, allowing 1d floodplain flow through its whole cross-section with a Manning-Strickler equation. However, in addition to being almost impossible to calibrate, the floodplain flow being extremely sensitive to the floodplain Manning coefficient because of its huge cross-section, the changes in terms of results were extremely negligible: less than 1 % of the water discharge in the main channel, well below expected model precision.

Another key assumption in SWAT-Amazon is the absence of backflow from the floodplain to the main channel during the rising limb. Observations in the Amazon basin support this, as low-gradient floodplain channels often experience tributary flow blockage, with the Tapajós River in Brazil being an extreme example. However, during deflooding, floodplain drainage dynamics differ from flooding processes. In SWAT-Amazon, these dynamics are governed by floodplain cross-section geometry. Introducing a Maillet-type law could improve recession representation, but discrepancies in timing appear mainly driven by rainfall variability, masking this effect. Thus, no modifications were made.

Floodplain sediment concentration during deflooding was assumed constant. Nevertheless, the unique monitoring of sediment concentration in a floodplain channel, performed at Lago Grande de Curuai (Brazilian Amazon) shows a decrease from ∼ 800 mg L⁻¹ during low-waters when the floodplain water flows back into the main channel to ∼ 40 mg L⁻¹ when the channel is under the influence of the Amazon River, with a mean concentration of 135 mg L⁻¹ (Moreira-Turcq et al., 2013). This value appears to be relatively close to that calibrated for Cbk (for the sand fraction only) at sub-basins 23, 22, and 21 (∼ 180 mg L⁻¹).

A dynamic law linking Cbk to water level could refine peak resuspension modelling but requires concentration data for calibration. Alternatively, Fagundes et al. (2023) treated the floodplain as a homogeneous sediment reservoir, balancing suspended fluxes with settling. However, this method applies only to clay and silt, excluding sand, and the settling law can be challenging to calibrate in the absence of observational data. Given floodplain water storage timescales (months), most fine sediments likely settle before being remobilized. Furthermore, as underscored in the introduction, sediments resuspended during recession can be centuries to millennia old, indicating long-term floodplain recycling rather than short-term remobilization.

Lastly, the distinction between Ebk and Ebed is partly supported by the calibration experiments. When bed erosion is activated (kbed>0), the model tends to generate rapid and abrupt peaks in simulated sand flux once the transport capacity exceeds the available sand load, whereas adjusting Cbk produces a smoother and more progressive increase in sand concentration that better reproduces the secondary peaks observed during the recession phase. The calibrated values of Kbed remain very small (Kbed≪1), suggesting that riverbed erosion contributes only weakly to the simulated suspended sand flux in the main stem, although a minor contribution cannot be fully excluded.

A GLUE-based uncertainty analysis was conducted over a Latin Hypercube ensemble of 2500 simulations (S12 in the Supplement), scoped to routing and sand transport parameters, conditional on the hydrological forcing, with physically informed priors (Sect. 6.1).

For discharge and sand flux, GLUE-weighted envelopes (5th–95th percentiles) remain narrow relative to the interannual amplitude, indicating that parameter uncertainty propagates weakly compared to the magnitude of hydrological variability and that the dominant dynamics are robustly constrained. Envelope width depends jointly on data availability and measurement quality, which act as coupled controls on parameter identifiability rather than independent sources of uncertainty. Floodplain water storage uncertainty in particular is strongly controlled by hf, emphasizing the need for accurate stage–discharge rating curves to constrain this parameter reliably (Sect. 6.2.3). Importantly, because the budget indicators (trapping and recycling fractions) depend primarily on hf, B, and db, which are well-constrained parameters (Sect. 6.1.3), they remain robust across the behavioural ensemble despite the wider prior ranges assigned to less identifiable floodplain parameters. This structural stability is confirmed by the GLUE-propagated budget envelopes (S11 and S12 in the Supplement).

Fine sediment fluxes achieve low bias and good performance metrics (Fig. 8b), yet exhibit wider uncertainty envelopes. This apparent performance is largely driven by the dominant seasonal signal rather than true predictive skill. Satellite retrieval errors cancel over long-term integration, masking short-term variability: the method reliably captures the mean seasonal cycle, but not its anomalies. Satellite radiometry is therefore robust at the climatological scale but not at the event scale.

Both conclusions rest on a common foundation: physically meaningful uncertainty quantification requires prior knowledge, which itself requires long-term, multi-variable observations, as provided by networks such as the CZO HyBAm. Uncertainty is therefore not only a modelling issue, but fundamentally a data-structure constraint. Model calibration and interpretation depend on sustained fieldwork and on a detailed understanding of measurement protocols and data limitations. As a result, rigorous data governance, FAIR practices, and strong research–operational links are essential for reproducibility and long-term continuity.

Water and sediment peak fluxes at the Andean outlet occur in February (Fig. 11), declining sharply from March to May as Andean precipitation decreases. The sediment flux at the basin outlet is strongly correlated with Andean discharge (Fig. 12c), confirming the Andes as the primary sediment source. This results from intense erosion along the Eastern Cordillera and Sub-Andean thrust belt, where high precipitation erodes Paleozoic and Mesozoic sedimentary sequences incorporated into Cenozoic deposits eastward as the orogenic prism advances through crustal shortening and thickening (McQuarrie, 2002a, b; Espurt et al., 2008; Gautheron et al., 2013; Pfiffner and Gonzalez, 2013). Additionally, the Central Andes' convex hypsometric profile (Montgomery et al., 2001) promotes deep fluvial incision.

At the lowland outlet, discharge peaks two months after the Andes (April–May) delayed primarily by alluvial plain storage, diffusive flood wave propagation, and runoff from the northern part of the basin, where seasonality, driven by the South American monsoon system (Garreaud et al., 2009), is less pronounced than in the south. This upstream-to-downstream flooding dynamic leads to progressive floodplain inundation (February–April, Fig. 11b, c) and gradual drainage (March–July). Floodplain backflow (Fig. 12a) drives sediment remobilization from April to August (Fig. 12b, d). This secondary sediment source accounts for ∼ 22 % of the total suspended sediment load at the basin outlet and can have significant impacts on river dynamics, with pronounced effects during major floods (e.g., 2012, Fig. 6c). Floodplain flows, comprising precipitation, river inflows and backflows, are more significant than the water entering the floodplain from the main river (Fig. 12a), reinforcing the assumption that floodplain waters are blocked by the main channel during the rising limb.

In the following analysis, the fine sediment load of the Pachitea Basin was estimated using a rating curve derived from the relationship between water discharge and measured fine sediment load, based on simulated flows. For the other Andean sub-basins (8, 9, 3), fine sediment load was regionalized according to the drainage area of their respective Andean catchments. The Ucayali floodplain consists of three geomorphological compartments (Fig. 2b), each characterized by distinct processes:

Lagarto – Pucallpa (Box 1, Fig. 2b): A major thrust fault system involving the crystalline basement exerts significant tectonic control over the narrow sedimentary basin wedged between the young Shira Mountains and the Fitzcarrald Arch (Espurt et al., 2008; Pfiffner and Gonzalez, 2013; Gautheron et al., 2013). Here, 36 % of the 420 [386, 454] × 10⁶ t yr⁻¹ of Andean sediment accumulates, with sand deposition (-100 [-83, -117] × 10⁶ t yr⁻¹ ) accounting for 66 %. This results in alluvial plain aggradation at 3.7 mm yr⁻¹ (9750 t km⁻² yr⁻¹), highlighting strong neotectonic control. In this compartment, water storage reaches 1.5 [1.1, 1.9] km³ in March on average, with flood durations (days with h≥hf) varying from 0 to 56 d in sub-basins 19 and 5, respectively. Consequently, daily-scale Andean flash floods superimposed on the annual monomodal flood wave are rapidly attenuated, while downstream, water storage increases as the floodplain expands. This process also dampens high-amplitude fine sediment peaks (-51 [-66, -35] × 10⁶ t yr⁻¹) through the activation of flood channels. These channels, remnants of the river's past mobility, further enhance the buffering effect. The sudden floodplain expansion at sub-basin 14 as the river course is released from the tectonic constraint imposed by the Shira Mountains reduces fine sediment loads (Fig. 11d), while high meander migration rates (Li et al., 2023) indicate active lateral erosion and deposition.

Pucallpa – Contaya Arch (Box 2, Fig. 2b). This section is a large piggyback basin located between the Andean wedge top and the Moa Divisor thrust fault system, with increased floodplain storage and longer flood durations (78–125 d). In March, it retains 10.4 [8.3, 12.5] km³ of water (∼ 20 % of discharge). Fine sediment load increases by +10 [+2, +17] × 10⁶ t yr⁻¹ due to Andean inputs, mainly from the Pachitea sub-basin, where one of the major Andean precipitation hotspots drives intense erosion rates (∼ 4,020 t km⁻² yr⁻¹, ∼ 1.5 mm yr⁻¹ ). Similarly, sands show a slight increase of +4 [+1, +7] × 10⁶ t yr⁻¹ after accounting for Andean lateral inputs, resulting in a total flux increase of +14 [+3, +24] × 10⁶ t yr⁻¹. Floodplain drainage and sediment resuspension gain importance in this segment.

Contaya Arch – Requena (Box 3, Fig. 2b). The river shifts northeastward into the extensive Marañón Foredeep, a tectonic funnel terminating at the Iquitos Arch. In April, water storage reaches 9.4 [7.7, 11.1] km³ . The maximum flood duration reaches 125 d. The flood impacts both fine sediments and sands (Fig. 11), however with different processes. Fine sediment deposition (-16 [-23, -10] × 106 t yr⁻¹, 7 % of the load) is driven by prolonged floodplain residence time, while sand sedimentation (-11 [-13, -9] × 10⁶ t yr⁻¹) results from energy dissipation when the floodplain is active. This process, never documented before in Amazonian rivers, captures approximately 14 % [10 %, 20 %] of the sand flux.

Overall basin budget. Integrating these three compartments, the Ucayali floodplain traps 165 [138, 192] × 10⁶ t yr⁻¹ of the 455 [410, 500] × 10⁶ t yr⁻¹ supplied by the Andes (∼ 36 %), with sand accounting for 65 % of total deposition. The downstream export to the Amazon River reaches 290 [235, 345] × 10⁶ t yr⁻¹ (26 % sand), identifying the Ucayali as the dominant sediment source among Andean foreland tributaries of the Amazon, exceeding the Marañón (∼ 144–173 × 10⁶ t yr⁻¹; Armijos et al., 2013) and the Beni (∼ 133 × 10⁶ t yr⁻¹; Vauchel et al., 2017). It also exhibits some of the highest specific erosion and deposition rates in the foreland. Net fine sediment loss across the basin amounts to 21 % of the Andean input, driven primarily by the dampening of Andean flash floods in Compartment 1. Floodplain recycling during the recession limb contributes approximately 22 % of the total suspended load at the basin outlet, acting as a secondary sediment source that partially offsets net deposition and sustains sediment export during the low-water season. Peak floodplain water storage reaches 19.1 [15.3, 22.9] km³ in March, equivalent to ∼ 38 % of total basin discharge for that month.

No significant trends were observed in flood duration, floodplain water volume, discharge extremes, or sediment fluxes in sub-basins 21, 22, and 23 for the simulation period. This suggests that the effects of global and regional changes at the Ucayali Basin scale are still negligible, contrasting with studies of the broader Amazon Basin (e.g. Marengo, 2004; Gloor et al., 2013; Li et al., 2020; Fleischmann et al., 2023). However, most of these studies focus on a very limited number of sites with long-term data. This highlights the relevance of the integrated approach proposed in this study for generating new long-term analyses of regional trends and providing a more detailed perspective on the Amazon's response to global changes. Espinoza et al. (2009) is the only study on Ucayali discharge evolution, based on inconsistent 1996–2005 data, now curated here. Figueiredo et al. (2024) and Polasky et al. (2025) identified mixed precipitation trends in the Ucayali Basin, respectively for the 1982–2023 and 1980–2022 period, which may explain the lack of clear trends here. Nevertheless, Guimberteau et al. (2013, 2017) project increased precipitation in the western Amazon by the century's end, potentially leading to greater inundations, sediment deposition, and floodplain recycling, which may alter river mobility, particularly during extreme events. This could impact bedload transport, leading to bed aggradation, which would facilitate overflow and then intensify exchanges.

The floodplain-controlled sand sedimentation process identified here, whereby inundation reduces main-channel transport capacity and captures ∼ 14 % of the sand flux at peak discharge, generating a secondary sediment peak during recession, has implications beyond the Ucayali Basin. Large tropical river systems draining tectonically active forelands onto extensive alluvial plains are likely to exhibit similar morphological conditions, suggesting that floodplain-driven decoupling of sand flux from water discharge may occur more broadly. The integrated approach developed here provides a framework to identify and quantify this process in other poorly monitored basins.