The variable geographic origin of the river flow is recognized to be a major factor driving the variability (Bustillo et al., 2010) of chemical signals. This is due to the heterogeneous distribution of soils, rocks, relief and vegetation cover on the basin. A very common illustration is provided by black-water and white-water rivers whose chemical properties differ considerably with respect to salinity, sediment load and mineralogical feature (Gibbs, 1967). Actually, the heterogeneous distribution of climate characteristics (rainfall, temperature, relative humidity, potential evapotranspiration, etc.) leads to nonsynchronized flood waves at the scale of the basin (Fig. 2). Usually, the propagation of rain waves is early in the southern part of the basin and delayed in the northern part. The time-lag between northern and southern areas is ca. three months. Consequently, the contribution of southern rivers (Madeira, Purus, and Juruá) arrives earlier than rivers draining northern areas (Negro, Japurá, Iça). Because northern and southern basins drain areas of very distinct characteristics, chemical signals impulsed in each area are very different: low nutrient concentrations, low TDS, low sediment load and high [DOC] for black-water rivers coming from north vs. high nutrient concentrations, high TDS, high sediment load and low [DOC] for white-water rivers coming from south. This generates a chemical dissymmetry between rising and falling water stages of the Amazon River.

Using the multiyear time-series chemical data set of Marchantaria (n = 108; period: 1983–1993; Devol et al., 1995) located on the Amazon River, the variability of chemical signals related to the variable contribution of three regional sources (Andes, major tributaries and local source) could be roughly appreciated for 24 parameters. The chemical characteristics of these three regional sources were established by means of multilinear regressions (Table 1). These indicate that a significant part of variability (0 to 50%) is associated with variable contributions of regional inputs. The contribution of the Andes imprints very characteristic chemical signals: sediment-laden waters and high concentrations of sediment-associated species (FPOM, CPOM), high TDS, high pH, high level of dissolved O₂, low CO₂ content, and high [NO₃⁻]. The tributaries impulse chemical signals exactly opposite to those impulsed by the Andes. Finally, the contribution of local sources leads to intermediate contents, except for SiO₂ (high content), phosphorus (dissolved and particulate, low contents) and alkalinity (low value).

At the scale of each subbasin, each local area exhibits a variable response-time depending on the soil characteristics (mainly depth and structure), morphology (slope, river network, and distance to the outlet) and climatic hazard. Taking explicitly into account this variability implies to implement a distributed hydrological model, coupled to GIS tools. Preliminary results from Victoria (2010) and Bustillo (2007), relying on the application of the macro-hydrological VIC model (Liang et al., 1994) on the whole Amazon basin, and forced by the climate dataset CRU05 (New et al., 1999), clearly outline the difficulty of the approach, mainly due to the poor precision of rainfall data and soil depth estimations. Thus, it was decided to take each major subbasin as a whole, and to consider primarily their discharge at the outlet, usually available. If not, the discharge may be assessed from satellite-derived water levels (TOPEX altimetry, Zakharova et al., 2006).

The heterogeneities within each subbasin, related to the variable contribution of hydrological sources (runoff components), is a major factor of flux and concentration variability. The rainfall water can take numerous pathways before reaching the streams: overland flow, shallow subsurface flow and groundwater. The conditions encountered by water along flowpaths control biogeochemical processes that determine stream water chemistry. The compositional seasonality of the stream water is intrinsically related to the variable contribution of hydrological sources to the river flow: for example, the flood events (i.e. related to surface quickflow) are responsible for up to 99% of suspended matter annual input to the Berre Lagoon, in France (Gouze et al., 2008). Several methods were attempted to identify end-members and to capture their characteristics. To achieve this purpose, hydrograph separation methods are commonly implemented, including chemical tracers (Mortatti, 1995; Tardy et al., 2005), isotopic tracers (Mortatti et al., 1997), digital recursive filtering (Eckhardt, 2005; Gonzales et al., 2009), spectral analysis (Spongberg, 2000), PCA-based methods (Chaves et al., 2006; Christophersen and Hooper, 1992; Hooper et al., 1990) and distributed hydrological modeling such as VIC 3-L (Victoria, 2010). Therefore, it is possible to determine at each time step the specific discharge of each runoff contributing to river flow. An iterative procedure was proposed by Bustillo (2005) and Bustillo et al. (2011) to determine the most likely solution, as a function of geochemical criteria (minimization of squared errors between observations and simulations, over a wide selection of chemical parameters) and hydrological criteria (adequacy with digital recursive filters). This methodology was successfully applied to the Amazon River and its main tributaries and allowed the identification of three hydrological sources: early direct runoff RS, delayed floodplain emptying RI and baseflow RB. The chemical contrasts between each runoff are unequivocal and for most of the parameters, the seasonal variations of chemical composition in the river water can be interpreted as the result of the mixture in variable proportions of constant composition reservoirs (Tardy et al., 2005; Tardy et al., 2009). In some cases, the fluxes are clearly nonconservative, because of in-stream diagenesis, lateral exchanges or sedimentation vs. remobilization patterns. The gases (CO₂, O₂), nutrients (NO₃⁻, NH₄⁺, HPO₄²⁻), coarse suspended sediments (sand fraction) and coarse particulate organic matter (POCC and PONC) do not exhibit a conservative behavior, and consequently, classical mixing models do not apply reliably (Bustillo et al., 2010). However, the compositional variations are strongly related to the hydrological conditions: river flow, direction of lateral exchanges between the main stream and the floodplains (Bustillo et al., 2010). The floodplain runoff, given by RI, constitutes an adjustment tool accounting for the nonconservativity of fluxes, and therefore, forecasting the seasonal and interannual variability of these nonconservative parameters is unproblematic. For the other considered parameters, at the scale of subbasins, the variable contribution of hydrological sources is the key factor driving compositional fluctuations in the river water (Tardy et al., 2009).

The long residence time in the stream network and the supply of unweathered sediments, in aquatic environments with high temperatures (28–30 °C) and oxidative conditions, promote the occurrence of biotic and abiotic processes that might alter substantially the chemical signals imprinted upstream, in soils and headrivers. As the fluxes are not perfectly conservative, end-member mixing models do not apply strictly, more particularly because the floodplains act as a natural fluvial filter of land-to-ocean fluxes (Meybeck & Vörösmarty, 2005). It is the case for major cations (Na⁺, K⁺, Ca²⁺, Mg²⁺) released by chemical weathering of incompletely weathered sediments deposited in floodplains for instance (Martinelli et al., 1993). Bioactive elements such as NO₃⁻, NH₄⁺, HPO₄²⁻, SiO₂ but also organic molecules and gases (O₂, CO₂, CH₄, N₂O) undergo important transformations along their course in streams: biological uptake (e.g. nutrient enrichment test in Calado lake from Setaro & Melack, 1984) and/or release (e.g. using ¹⁸O/¹⁶O in dissolved O₂ as a tracer of respiration and photosynthesis: Quay et al., 1995), biomass production (Wissmar et al., 1981), sorption/desorption (e.g. mixing zone of the Negro-Solimões confluence: Moreira-Turcq et al., 2003b), outgassing (Devol et al., 1988; Richey et al., 1988), photooxidation (Amon & Benner, 1996), etc. The situation is comparable for sediments and associated POC, definitely discharge-dependent for the coarse fraction CSS and POCC while FSS and POCF dynamics in the lower Amazon are unambiguously driven by longitudinal (floodwave propagation and backwater effect) and lateral hydraulic gradients (Meade et al., 1985).

Because the chemical composition of the Amazon River water cannot directly be inferred from the simple mixture of waters provided by major tributaries, an attempt is thus made to account for those processes arising in the river. Two methods are investigated: (1) description of baseflow composition changes as a function of the hydrobiological regime (Bustillo et al., 2010) and (2) statistical-regressive modelling taking into account the hydrological balance of floodplains. Considering the case of the Amazonian floodplain lake called “Lago de Curuai” located close to Óbidos, it was shown (Bonnet et al., 2008; Maurice-Bourgoin et al., 2007) from in situ and satellite data acquired between 1997 and 2003, that the Amazon River dominated the inputs of water to the flooded area, accounting for about 77% of the annual total input; rainfall and runoff account for about 9 and 10%, respectively, while seepage from the groundwater system accounts for only 4%. It results that, as a first approximation, the water budget of the floodplains can reasonably be assessed from the variations of the river flow. Moreover, when comparing the input by major tributaries with the output recorded at Óbidos, Moreira-Turcq et al. (2003a) found that the amount of organic carbon increased (about 4 Tg C yr⁻¹ i.e. + 12%), suggesting that important sources of autochthonous organic carbon may exist in the lower reaches of the Amazon River. These inputs were attributed to adjacent floodplain lakes, with intermittent supply related to the pattern of river-floodplain water exchanges (Bustillo et al., 2010).

The reconstitution of hydrograph separation aims to revisit the interpretation given to the fluctuations of the river flow by the calibration of specific hydrological relationships for each reservoir RS, RI and RB. The composition of each individual runoff: RS (direct runoff), RI (delayed floodplain emptying) and RB (baseflow) is assessed by a linear programming method, according to the protocol detailed by Bustillo (2005). It fundamentally consists in optimizing the discharges QRS, QRI and QRB so that (1) the simulated concentrations of 12 representative parameters, obtained by multilinear regression, minimize the mean relative squared error between modelled and observed data and that (2) QRS (t) + QRI (t) + QRB (t) = Qt (t). The reference values ascribed to QRS, QRI and QRB are those obtained by multilinear regression, with optimised sets of QRS (j,t), QRI (j,t) and QRB (j,t) for each station (index j). Several methods are tested to reconstitute QRS (j,t), QRI (j,t) and QRB (j,t) as a function of the only available parameter measured continuously, namely the river discharge Qt (j,t), estimated by means of frequently calibrated stage-discharge relationships.

4.1.1 Linking Qk (j,t) and Qt (j,t + n)

4.1.2 Digital filtering methods

At the confluence of large rivers, for which the draining areas contributing to the total discharge are submitted to phase-shifted hydroclimatic regimes, the procedures which consist in reconstituting the components of river flow by analysing only the total river flow tends to fail. Considering the 11 sampling stations located along the Amazon main stream, the composition of the river water at the station x and at time t, noted Cix(t), was estimated by integrating inputs from the major tributaries:

To assess the relative importance of autotrophy vs. heterotrophy on the water chemistry, the hydrobiological index IBIO = [O₂]–[CO₂] was defined (Bustillo et al., 2010). This synthetic index, corresponding to the difference between dissolved gas composition of O₂ and CO₂ in the water, enabled to assess the metabolic balance between autotrophy and heterotrophy in the fluvial system: IBIO > 0 indicates that photosynthetical patterns predominate and conversely, IBIO < 0 indicates that heterotrophy is the dominant hydrobiological pattern. Biologically-mediated processes were recognized to modify more specifically the composition of the baseflow. We assumed that the chemical composition of the baseflow RB fluctuated in accordance with IBIO so that:

[C]RB (j;t) = [Co]RB (j) + KBIO (j) × IBIO (j;t) (11)

where [Co]RB is the concentration in the baseflow before transiting in the lower fluvial reaches. KBIO ≈ 0 for chemical species exhibiting a nearly conservative behaviour such as Na⁺, Cl⁻, etc… and KBIO ≠ 0 for bioactive elements such as NO₃⁻, CO₂, etc. Therefore, it is necessary to evaluate IBIO (j;t) variations to improve the chemical characterization of RB and subsequently the assessment of and biogeochemical budgets. The evaluation of IBIO (j;t) requires the concomitant evaluation of [CO₂] and [O₂]. In most of the cases, the dissolved gas composition of river water depends on the water turbidity (appreciated by FSS), on the river flow (Qt) and on the discharge of the floodplains, tracked by ΔQt/Qt = (Qt_(t+1)–Qt_(t-1))/Qt_(t) involving the discharge of the month Qt_(t) and those of the next month Qt_(t+1) and last month Qt_(t-1). Actually, the water turbidity controls the yield of aquatic photosynthesis: sediment-laden waters promote IBIO < 0. Moreover, the river flow determines the extension of flooded areas where the organic matter decay is promoted. If the river flow is low, the residence time of water in streams tends to lengthen while organic substrate is almost nil so that photosynthesis prevails: IBIO > 0. Finally, the intermittent filling/emptying of the floodplains (appreciated by ΔQt/Qt), where heterotrophic regime prevails, tends to modify the dissolved gas composition of streams, by releasing CO₂ and removing O₂. Because of the well-known interrelations of these factors, we chose to calibrate the following equations:

This correction applies to the chemical composition of the Amazon River main reach. Due to the impossibility of determining the default of hydrological and chemical budget within the major tributaries (where there is only one monitoring station at the outlet), the procedure exposed here cannot be applied to them. The variation of chemical composition, noted ΔCijk, between theoretical (calc) and observed (obs) values is related to two factors: the river flow at the x^th sampling station: Qt (xk), and the default of water balance in the floodplains WBF_xk.

The default of water balance WBF implicitly takes into account lateral exchanges between floodplains and alluvial aquifers, on the one side, and the river main channel on the other side. It is particularly the case for POCC and CSS which are sequentially released to the river when floodplains dry up (Bustillo et al., 2010), i.e. when WBF_x,k > 0.

For each parameter, four sources of variability were identified: (1) geographic source contribution, (2) hydrological pathway, (3) in-stream biogeochemical processes and (4) hydrological budget of the floodplains. Here, their respective contribution to the total variability of chemical signals in the Amazon River water is assessed. First, we have simulated the theoretical composition of the Amazon River at Manacapurú, São Jose da Amatari, Paurá and Óbidos assuming that the variable geographic source contribution is the only source of variation, called interbasin variability. Hence, as a first approximation, the chemical variability within subbasins, called intrabasin variability, is neglected. We suppose that the chemical composition of tributaries is constant and that the chemical variability in the Amazon River depends on the variable contribution of these tributaries to the Amazon flow. Simulated concentrations are compared to observed ones. The Explained Variance coefficient (EVC) (e.g. Franchini et al., 1996) was calculated to evaluate the goodness of fit and to assess the contribution of each source of variations on the compositional chemical variability in the Amazon River:

Step 1. The compositional variations associated to the interbasin variability (noted Cˆi,x,t*) are simulated by discharge-weighting of the contributions of the n main tributaries located upstream from the sampling station, considering that their individual concentrations are temporally constant:

Step 2. The compositional variations associated with the intrabasin variability (noted Cˆi,x,t**) are simulated by discharge-weighting the contributions of the n main tributaries located upstream from the sampling station, considering that their individual concentrations are temporally variable:

Step 3. The variability explained by in-stream biogeochemical processes is defined here as the gain of the Explained Variance Coefficient associated with the incorporation of IBIO _x,k = [O₂]_x,k–[CO₂]_x,k as a covariate:

Step 4. Finally, the variability explained by the water budget of the Amazonian hydrosystem (floodplains and main reach) is defined in this study as the gain of EVC associated to the incorporation of two hydrological descriptors, namely WBF and Qt, to describe the compositional variations of the chemical parameter i at the sampling station x at time t:

It must be kept in mind that the calculations are based on observed concentrations and discharges in the tributaries and not by the outputs of mixing models. As a result, the so-called intrabasin variability (step 2) constitutes an upper limit of that which would be obtained on the basis of hydrograph separation methods. Likewise, I_BIO (step 3) is calculated from observed [O₂] and [CO₂] in the main stream, meaning that the variability attributed to in-stream biogeochemical processes is an upper limit of that which would be obtained on the basis of estimated I_BIO.

The comparison of the methods relies on their individual fitting capability to estimate QRS (j,t), QRI (j,t) and QRB (j,t) obtained by the reference hydrograph separation. The correlation coefficients R² of each procedure are presented (Table 2). The indices n (–3 ≤ n ≤ + 3) correspond to the time-lag (n given in months) between Qt and Qk used to perform the linear calibration. The values which are underlined stand for maximum R² for this kind of procedure. It is remarkable that maximum R² is obtained for n = 2 in the case of RS, n = 2 in the case of RI and n = 0 in the case of RB. The largest phase shifts are obtained on climate contrasted basins such as Purus, Jurua and Madeira river basins. The results obtained by linear combination of Qt (t + n) deliver most of the time very good correlation coefficients (see column entitled LC). Auto-recursive methods are also quite efficient, and more particularly the mixed approach F3 consisting in adjusting (i) RI and RB during falling water, and (ii) RS and RB during rising water. The goodness of fits provides, a posteriori, a validation of end-member mixing models whose outcomes appear consistent from an hydrological point of view, except for the Japura, Iça and Negro Rivers for which RB and RS are poorly explained.

The variations of the stream waters chemical composition along a hydrological cycle is sometimes modelled as a dilution of some source water (baseflow) by varying volume of some other water type (quickflow). This approach relying on the mixing of two hydrological sources is sometimes efficient for small rivers, but it appears rarely valid for large rivers (Mortatti, 1995). The methodology applied by Tardy et al. (2005) enables to identify three hydrological sources. It relies on the use of chemical tracers, chosen to be fine suspended sediments (FSS) and Na⁺. These two parameters were supposed to exhibit a nearly conservative behavior in the streams. Actually, this assumption is not completely valid for FSS which undergo deposition and remobilization patterns. In the case of black-water rivers, most of FSS undergoes sedimentation so that surface runoff signal is strongly altered. It is probably more appropriate to choose DOC as tracer instead of FSS because quickflow (surface runoff and subsurface flow) are highly enriched in DOC compared to the slower and deeper flowpaths that contribute to baseflow. This justifies to estimate QRS (j,t), QRI (j,t) and QRB (j,t) by means of linear programming methods implying more than two tracers (e.g. Bustillo, 2005), or by means of PCA-based methods (e.g. Hooper et al., 1990) in order to minimize the impact of analytic errors and to improve the robustness of the assessments.

It should also be kept in mind that large river basins are spatially organized. Consequently, whatever the actual flowpath contributions to river flow, the variable regional contribution promotes important compositional fluctuations. As a matter of fact, the chemical tracing does not provide rigorously a picture of the soil vertical organization for the large river basins spatially organized (i.e. with mountains, piedmonts, low plains). Instead, it delivers an insight on the basin regional organization: (1) RS tracks the contribution of upstream areas, providing sediment-laden water; (2) RB tracks the contribution of subterranean water characterized by high TDS (where chemical weathering is the most active) and low sediment load; (3) RI, whose expression is delayed in time, tracks the contribution of saturated areas (fluvial corridors, floodplains), mainly located downstream, and usually provides low sediment load and low TDS.

RI is thus assimilated to a storage reservoir bordering the river network, fed concomitantly by RS and RB during rising water, and releasing water when local hydraulic conditions are favorable, i.e. mainly during falling water while adjacent water reserves are still high as river water level decreases.

Fig. 3 synthetizes the parameters of multilinear equations relating CO₂ and O₂ to the most significant physical drivers of river ecology: the water turbidity (approached by FSS), the river flow (Qt) and the contribution of the floodplains (tracked by ΔQt/Qt) to the streamflow. This analysis indicates that [CO₂] is very low when the river flow is low to moderate and when waters are free of turbidity (FSS = 0); conversely, under these conditions, [O₂] is high, indicating that aquatic photosynthesis prevails. Since the rise of turbidity indicates increasing contribution of surface runoff to the river flow, and seeing that this surface runoff drains upper layers of soils which constitute oxygenated environments, it can be concluded that the influence of turbidity is not univocal. That is the reason why, in some cases, the decrease of water transparency is more than compensated by the input of quickflow R_S, close to air saturation (α₂ > 0 for Rios Iça, Jutaí, Jurua, Japura, Madeira) and in other cases, not compensated (α₂ < 0) due plausibly to high oxygen demand from transported carbon: high DOC for the Negro and the Solimões Rivers, high POCF for Solimões and Purus Rivers. The influence of turbidity on O₂ contents is not significant for the sampling stations located along the Amazon River studied reach (−0.5 < α₂ < 0.1).

The relationships between the dissolved gas composition of water and the discharge appear much easier to decipher. [CO₂] increases (β₁ > 0) with the river flow (Qt) while [O₂] decreases (β₂ < 0), indicating that the yield of aquatic photosynthesis drops when the discharge rises. Moreover, we notice that dissolved gas composition of river water is closely linked to the rate of variation of discharge. For a given discharge, [CO₂] is higher during falling water (γ₁ < 0) than during rising water. Conversely, [O₂] is higher during rising water (γ1 > 0). Actually, the discharge of the floodplains and wetlands which border rivers drain water enriched in CO₂ and depleted in O₂ because organic carbon is specifically and intensively mineralized in stream corridors (Mayorga et al., 2005), which mainly feed river flow during falling water (Richey et al., 1989; Alsdorf et al., 2010).

The results are discussed in light of synthetic 2-D diagrams (Fig. 4). These diagrams provide a qualitative insight on river diagenesis as a function of the discharge time-series and the default of discharge balance WBF. Four parameters were selected for illustrative purpose: Na⁺, DOC, FSS and O₂. Based on a calibrated set of parameters (α_ix, β_ix, γ_ix and δ_ix), the mean annual fluctuations at the station of Óbidos were represented. These diagrams show that seasonal concentration patterns of these parameters are not only influenced by the river flow (x-axis) but also by the contribution of the floodplain to the river flow. DOC and O₂ exhibit comparable patterns marked by increasing concentration (for a same discharge value) for a floodplain emptying pattern. This suggests that floodplains are sources of DOC and O₂ for the fluvial system. However, most of the variability is associated with dilution patterns for rising discharges. It is more contrasted for Na⁺ whose surface response might secondarily be influenced by the timing of early alluvial groundwater inputs. Conversely, the fluctuations of silt-clay suspended sediments (FSS) appear to be equally influenced by the discharge and by the water balance of the floodplains. FSS exhibit an unexpected hysteretic C-Q shape marked by higher concentrations during floodplain emptying, due to the flushing effect of the Negro River (Dunne et al., 1998) whose flow peak is delayed with respect to those of the other tributaries.

Fig. 5 presents the values of the Explained Variance coefficient (EVC) calculated at four stations (sorted from upstream to downstream) located on the Amazon River main stream: Manacapurú, São Jose da Amatari, Paurá and Óbidos. The values obtained for EVC and presented in Fig. 5 are cumulative. For each couple [station; parameter], four EVC values are established, corresponding to the successive incorporation of tested factors in the modelling process. EVC commonly exhibit values > 75%.

EVC related to variable geographic source contribution is most of the time < 40%. Conversely, it is worth noting that variable geographic source contribution explains more than 80% of the total variability for Ca²⁺ and HCO₃⁻, and more than 50% of Mg²⁺ and Cl⁻. The interbasin variability of these parameters is much greater than the intrabasin variability (mainly related to hydrological source) and moreover, these chemical species exhibit a nearly conservative behavior in the river. Complementarily, it should be mentioned that most parameters exhibit high determination coefficient and low Nash-Sutcliffe coefficients (data not shown). It is typically the case for δ¹³C (DIC), CO₂, O₂, pH, NO₃⁻ for example. Here, the variable geographic source impulses a chemical signal which tends to be amplified by hydrological source and in-stream diagenesis (Bustillo et al., 2010), because the biogeochemical processes arising downstream are prolonging those operating upstream from the confluence with the Amazon River main stream.

The variability related to hydrological pathways, called intrabasin variability, is appreciated by comparing the simulated (tributaries-derived) and observed composition of the Amazon River water. Its contribution to the overall variability of concentrations is very significant. In some cases, the simulated data variance is larger than the observed one, indicating that the chemical signals imprinted upstream tend to be buffered downstream. It is the case for Ca²⁺, HCO₃⁻, SO₄²⁻, DOC, DIC, FSS, CSS, POCF, POCC, δ¹³C (POCC), PONF and PONC.

Then, incorporating the influence of in-stream biogeochemical processes enables to improve the amount of explained variability with respect to chemical signals in the Amazon River water. Most of EVC are greater than 90%, except for SO₄²⁻, CSS, POCC, δ¹³C (POCC), PONC and POCC/PONC. Actually, the model of in-stream diagenesis integrates interbasin and intrabasin variability, in addition to the variability impulsed by in-stream processes. As expected, the part of variance explained by the models typically rises as we add a new factor. It is therefore possible to identify successively the variation due to each factor. For example, taking the case of Na⁺ at Manacapurú, Fig. 5 indicates that 4% of the variability is due to interbasin variability, 69% (0.69 = 0.73–0.04) is related to intrabasin variability and 23% (0.96–0.73) to fluvial processes. In peculiar and limited cases, (inter + intrabasin) variability is lower than the interbasin one (e.g. Ca²⁺ at Paurá), probably due to interaction effects that jumble out the chemical signals imprinted by geographic source variability. Among the two river factors influencing chemical signals of the Amazon River, namely hydrobiological processes and floodplain water balance, the latter is clearly the most impacting. It might be due to the fact that the seasonally-structured hydrological response of the floodplains impulses to a large extent [CO₂] and [O₂] variations in the main channel; since the water coming from the floodplains is typically depleted in O₂ and concentrated in CO₂; the water balance of the floodplains implicitly determines the imprint of hydrobiological processes arising there. Hence, the variability ascribed to the water balance of the floodplains integrates also a part of variability related to hydrobiological processes. The case of Na⁺ deserves to be discussed more thoroughly. Indeed, half of the increase observed on EVC (compared to inter + intra variability), ascribed to river processes, is related to I_BIO while the release vs. sequestration of Na⁺ is recognized to be fundamentally an abiotic process. This gain observed on EVC unduly attributed to I_BIO is actually due to the fact that the water budget of the floodplains (WBF) influences simultaneously I_BIO and [Na⁺]. Consequently, the increase of explained variability related to the addition of I_BIO as a potential source of variation should not be interpreted in terms of causality relationships between I_BIO and [Na⁺]. The actual driving factor is the WBF, as shown by Bustillo et al. (2010): the floodplains are sites where the chemical alteration of the unweathered deposited sediments is very active, so rising contribution of floodplains to the overall water budget of the Amazon River promotes thus increasing [Na⁺] in the main channel.

The contribution to the total variance of (1) geographic source variation (interbasin), (2) hydrological source variation (intrabasin) and (3) river processes, differs considerably from a parameter to another. For Ca²⁺ and HCO₃⁻, most of the variance is related to geographic source (60% to 93%); while the hydrologic source and the river processes have little impact on these parameters. Although less marked, comparable trends are observed for DOC, Mg²⁺, FSS, POCF, PONF and O₂. The case of δ¹⁸O deserves to be mentioned because most of the variance is attributed to a single factor, i.e. hydrological source (> 97%), indicating that interbasin variability and river diagenesis do not impact significantly δ¹⁸O signals in the Amazon River water. River processes seem to have a higher impact (10–30% of total variance) on [Na⁺], [Cl⁻], pH, [NO₃⁻], [HPO₄²⁻], [CO₂], POCF/PONF, DOC/DON, DON, δ¹³C (DIC), δ¹³C (POCF), δ¹³C (POCC), CSS and POCC. In turn, for CSS and POCC for example, the combination of these three sources of variations does not explain the variability of the chemical signals, attributed to that of the river surface slope (e.g. on the Madeira River, main provider of suspended sediments: Martinelli et al., 1988; on the Solimões River at Marchantaria: Devol et al., 1995) which might vary considerably due to the concomitance of large stage variations (up to 10 m) and very shallow thalweg slope (2–4 cm/km) that promote backwater effects (Meade et al., 1991).