Environmental Controls on the Riverine Export of Dissolved Black Carbon

© 2019 American Geophysical Union. All rights reserved. Jones Matthew William (Orcid ID: 0000-0003-3480-7980) Aragão Luiz E.O.C. (Orcid ID: 0000-0002-4134-6708) Dittmar Thorsten (Orcid ID: 0000-0002-3462-0107) Johnson Ben T. (Orcid ID: 0000-0003-3334-9295) Marques Jomar Silva (Orcid ID: 0000-0002-6722-3693) Niggemann Jutta (Orcid ID: 0000-0003-1237-9233) Quine Timothy (Orcid ID: 0000-0002-5143-5157)


Introduction
The riverine export of dissolved organic carbon (DOC) is one of the major fluxes of carbon across the land-to-ocean aquatic continuum . It is estimated that 208 ± 28 Tg of DOC year −1 is exported by rivers to the global oceans, predominantly from stocks of soil organic carbon (SOC) in their drainage catchments, with 62% of this export occurring in tropical rivers (Dai et al., 2012). The export of DOC by global rivers is an important process within the global carbon cycle, with global fluxes equating to ~7% of the net land sink for atmospheric CO2 (Le Quéré et al., 2018). On the order of 10% of this DOC is in the form of dissolved black carbon (DBC), which is composed of polycondensed aromatic molecules produced by the incomplete combustion of organic matter in biomass and fossil fuels (Jaffé et al., 2013;Wagner et al., 2018). The storage of terrigenous DOC in the global oceans regulates atmospheric concentrations of CO2 (Bauer et al., 2013;Bianchi, 2011) and DBC is a particularly pertinent fraction of this DOC because its chemical properties make it intrinsically resistant to degradation (Tranvik, 2018;Wagner et al., 2018). The recent realisation that ignoring lateral fluxes of carbon from terrestrial to marine environments results in nontrivial errors in terrestrial carbon accounting, combined with evidence for the anthropogenic perturbation of the DOC export, has triggered a renewed focus on identifying the environmental factors that control its export, character and fate (Battin et al., 2009;Cole et al., 2007;Raymond et al., 2016;Regnier et al., 2013). As part of this agenda it will be pivotal to develop a mechanistic understanding of the environmental factors that control the export of DBC across the land-to-ocean aquatic continuum (Coppola et al., 2018;Dittmar & Stubbins, 2014;Wagner et al., 2018).
The DBC exported by rivers derives from stocks of charcoal stored in the soils of river catchments (Dittmar et al., 2012a;Jaffé et al., 2013;Wagner et al., 2018) as well as from aerosol (soot) deposited from the atmosphere to river catchments (Jones et al., 2017;Wang et al., 2016). BC is a product of biomass and fossil fuel combustion and is distinguished from bulk OC by its condensed aromatic structure and low functionality (Dittmar, 2008;Masiello, 2004). These properties make BC highly resistant to chemical and biological decomposition in environmental matrices, which promote its accumulation in soils, sediments and oceanic DOC (Bird et al., 2015;Schmidt & Noack, 2000). If biomass carbon stocks are allowed to recover following a fire, the BC produced by biomass burning represents an additional terrestrial store of carbon (i.e. a sequestration sink for CO2) owing to its longevity in these stores relative to non-combusted organic carbon (Santín et al., 2015). A long-term net sink for atmospheric CO2 develops if this sink is not offset by CO2 evolution from legacy BC stocks created in previous millennia (Landry & Matthews, 2017). The transfer of BC from land to ocean is an important constraint on the balance of these CO2 sinks and sources because this extends its residence time from decades-to-centuries in soils (Kuzyakov et al., 2014;Singh et al., 2012) to centuries-tomillennia in oceanic stores (Coppola & Druffel, 2016;Coppola et al., 2014).
Over the past three decades, considerable progress has been made in understanding the factors that control the rates of DOC export from catchments to river channels. One element of this progress has been the identification of correlations between the environmental characteristics of river catchments and the observed concentrations of DOC in channels draining those catchments. Studies have been conducted at catchment scales ranging from relatively small headwater channels (Ågren et al., 2007;Aitkenhead et al., 1999;Clair et al., 1994;Dillon & Molot, 1997;Eckhardt & Moore, 1990;Frost et al., 2006;Gergel et al., 1999;Mulholland, 1997;Rasmussen et al., 1989;Wilson & Xenopoulos, 2008), to regionally and globally significant rivers (Aitkenhead & McDowell, 2000;Hope et al., 1997;Huang et al., 2012;Ludwig et al., 1996;Mattsson et al., 2005). These studies show that a small number of dominant factors drive the variability in riverine DOC concentration and fluxes. Predictors that typically explain significant portions of this variability include soil organic carbon (SOC) stocks, soil moisture, mean catchment slope, wetland cover and antecedent precipitation (Aitkenhead-Peterson et al., 2003;Harrison et al., 2005;Mulholland, 2003). The first of these factors represents the major stock of organic carbon that is available for mobilisation in the dissolved phase, while the latter factors relate principally to hydrological controls on the rate at which this organic carbon is accessed and mobilised (Neff & Asner, 2001;Rasmussen et al., 1989).
Statistical models fitted to empirical data have facilitated the construction of process-based numerical models of DOC export by constraining the role of environmental conditions in the export of DOC from catchments (Lauerwald et al., 2017;Nakhavali et al., 2017;Neff & Asner, 2001).
The studies discussed above have consistently highlighted the role of hydrology in determining rates of DOC export from SOC stocks in river catchments. As SOC is disproportionately stored in the uppermost soil horizons, the contact time between drainage water and SOC is subject to hydrological controls that determine the depth of the water table. Soil moisture can be viewed as a proxy for the depth of the water table, with higher moisture levels indicating that a larger component of the drainage flow path is in contact with the upper organic soil horizons (Inamdar et al., 2011;Lambert et al., 2011;Singh et al., 2014). Consequently, increases in DOC export by rivers are widely observed during wet seasons (Lambert et al., 2011;Singh et al., 2014) and isolated rainfall or storm events (Inamdar et al., 2011;Raymond & Saiers, 2010;Raymond et al., 2016;Stanley et al., 2012;Vidon et al., 2008;Yoon & Raymond, 2012). Soil moisture tracks the seasonal balance of precipitation and evapotranspiration plus discharge, thus representing the seasonal hydrology of a catchment. Catchment slope is also an indicator of the contact time between water and soil organic matter because the geometry of catchments determines the depth soils and the rate at which precipitation is conveyed to channels (Aitkenhead & McDowell, 2000;Dillon & Molot, 1997;Ludwig et al., 1996;Mattsson et al., 2009;Rasmussen et al., 1989;Sobek et al., 2007;Sutfin et al., 2016): steep catchments typically exhibit thin organic soil horizons and rapid rates of water transfer to drainage channels, thus restricting interaction between water and soil organic matter and moderating rates of DOC export, whereas; catchments with low slopes promote slow water drainage and the accumulation of thick organic horizons, thus enhancing contact time between water and soil organic matter and promoting DOC export.
Globally, a strong linear relationship has been identified between the concentrations of DBC and DOC in river channels, which suggests that DBC and DOC are exported from catchments in a coupled manner across the gradients in environmental conditions that occur at this scale (Jaffé et al., 2013;Wagner et al., 2018). DBC export fluxes are thus usually considered to be coupled to those of DOC export, with global-scale fluxes of DBC predicted by a simple linear relationship with equivalent estimates of DOC export (Jaffé et al., 2013). A potential explanation for this close relationship is that the process of DBC mobilisation from soil BC stocks is subject to the same environmental controls as DOC mobilisation from SOC stocks, including the hydrological factors discussed above. Nonetheless, there has not yet been a comprehensive analysis of the variability in riverine fluxes of DBC across the environmental gradients that are known to control rates of DOC export.
In the current study, we simultaneously assessed how rates of DOC and DBC export by rivers are controlled by the environmental conditions in South American tropical river catchments.
Tropical rivers have particular relevance to global DBC export because they contribute 62% of the global DOC export flux (Dai et al., 2012) and because around 90% of global burned area occurs in the tropics (Chen et al., 2017;Giglio et al., 2013). We fitted statistical models which robustly predicted DOC and DBC concentrations using hydrological, topographical, soil and climate variables, as well as upstream stocks of SOC and SBC, respectively, in channels whose catchments show significant diversity in these factors. Factors were chosen specifically because they are known to influence riverine DOC concentrations (Aitkenhead-Peterson et al., 2003;Mulholland, 2003) or because they have been implicated as processes governing the dynamics of DOC and DBC in soil (Kaiser & Kalbitz, 2012;Kuzyakov & Blagodatskaya, 2015;Kuzyakov et al., 2014).
We also investigated the role of aerosol BC deposition to river catchments as a driver of variability in DBC export from river catchments. Aerosol BC has traditionally been considered a negligible source of riverine DBC on the basis that it contributes less than 30% towards the total global production flux for BC (Bird et al., 2015). Nonetheless, a number of recent studies have challenged this view by demonstrating that aerosol BC makes contributions on the order of 5-25% of the total riverine DBC load in some high latitude (Ding et al., 2015;Stubbins et al., 2012a), temperate (Wang et al., 2016) and tropical (Jones et al., 2017) catchments. Further evidence is required in order to validate the contribution of BC aerosol to riverine DBC at regional and global scales (Bao et al., 2019;Wagner et al., 2018). The deposition rate of aerosol BC relates to regional rates of emission from several sources (e.g. industry, deforestation, wildfires and agriculture) combined with wind patterns and the efficiency of wet and dry deposition processes (Bao et al., 2017;Bond et al., 2013;Jurado et al., 2008). Consequently, the spatial distribution of aerosol BC in river catchments does not align with that of soil BC stocks. This study provided an opportunity to assess the influence of aerosol BC deposition on riverine loads of DBC at the continental scale.
We fitted generalised linear models (GLMs) with land cover; soil properties; recent weather and soil moisture conditions; stocks of SOC and SBC, and; aerosol BC in river catchments as predictors of DBC and DOC concentrations measured in Brazilian rivers between 2013 and 2016. The statistical power of the models is high owing to the large sample size of the reference dataset (192 samples) and the distribution of sampled channels across tropical forest and savannah biomes spanning a latitudinal range of over 20°. These catchments of the sampled rivers displayed significant gradients in the predictor variables considered in this study. Our analyses allowed us to explain why the concentrations of DBC and DOC from river catchments are generally coupled in river systems, while we also identify the factors that weaken this relationship.

Sampling Campaigns
The sampling locations were distributed within 6 study regions (figure 1), three within the tropical forest biome (AF1, AF2 & AMZ1) and three in the Cerrado (C1-3). AF1 and AF2 are situated within the domain of the Atlantic Forest, which stretches ~500 km inland from the Atlantic coastline of Brazil. Over the past two centuries the native Atlantic Forest has been pervasively cleared by slash-and-burn deforestation and today just 7-8% of the original 1.5 million km 2 of forest remains (Ribeiro et al., 2009). Study region AMZ1 is located in the 'arc of deforestation' in Southern Amazonia, where the forest frontier has moved progressively northwards over the past 50 years and fire has been a major tool for vegetation clearance (Fearnside et al., 2009;Malhi et al., 2008;Tyukavina et al., 2017). All three tropical forest study regions feature expansive agriculture with interspersed small forest fragments, although some larger forest fragments remain in national parks and private reserves (supplementary text S1). The samples collected in AF1, AF2, and AMZ1 represented a variety of headwater channels, tributaries and regionally significant river channels whose catchment areas ranged from 35 km 2 to 85,600 km 2 (table S1).
Study regions C1, C2 and C3 are located in the Cerrado, which dominates the natural land cover of the Brazilian interior. The biome is composed of several vegetation physiognomies ranging from low-lying, grass-dominated campo sujo and campo Cerrado to the increasingly shrub-dominated Cerrado sensu stricto, Cerrado denso and dry tropical forest (Cerradão) (Klink et al., 1995). Natural Cerrado vegetation once covered over 2 million km 2 of Brazil, around 25% of its territory, but today only 46% of this area remains following progressive disturbance over the past 50 years (Salazar et al., 2015). Since 1970 the biome has experienced widespread clearing by fire and conversion to planted pasture and irrigated cash crops (Klink & Machado, 2005;Oliveira & Marquis, 2002;Pivello, 2011). The Cerrado study regions featured a typical degraded Cerrado landscape; agricultural land uses are interspersed with fragmented sections of natural dry forest, shrubland and grassland. Natural cover is concentrated in national and state parks, and is otherwise restricted to drainage channels and gullies where the terrain is unfavourable for agriculture. The samples collected in AF1, AF2, and AMZ1 represented headwater channels and tributaries to major rivers, with catchment areas ranging from 10 km 2 to 1,300 km 2 . The samples collected in C1, C2, and C3 represented a variety of headwater channels and tributaries whose catchment area ranged from 11 km 2 to 1,300 km 2 (table S1). For further information regarding the regions studied, the reader is referred to supplementary text S1 (Fearnside et al., 2009;Klink & Machado, 2005;Malhi et al., 2008;Oliveira & Marquis, 2002;Pivello, 2011;Ribeiro et al., 2009;Salazar et al., 2015;Tyukavina et al., 2017).
We completed the sampling campaigns in AF1, C1, C2, C3 and AMZ1 between April and May 2016 during the transition from the wet season to the dry season. Sampling campaigns in AF2 were completed by Marques et al. (2017) during three periods with differing hydrological conditions, specifically in January 2013 (wet season), August 2013 (dry season) and February 2014 (wet season). River water discharge rates varied by a factor of 4 across the sampling campaigns at the most downstream gauge in AF2 and were 1,875, 478 and 719 m 3 s −1 , respectively (Marques et al., 2017). Our reference dataset thus included measurements from samples that represented both a large spatial domain and a range of hydrological conditions. © 2019 American Geophysical Union. All rights reserved.

Sampling Design
The aim of the sampling campaigns was to include catchments displaying wide variability in all explanatory variables included in the analysis. To achieve this, the sampling regions were positioned in multiple ecoregions of Brazil across which soil carbon stocks and climate were known to differ. Further, the spatial coverage of the study regions ensured that they were spread across a large gradient of aerosol deposition rates. Variability in land cover, soil type and clay content was achieved by sampling in many catchments within each study region, across which sufficient variability in these parameters was present. To represent variability in soil moisture and rainfall we incorporated data from samples collected over multiple seasons and a large spatial domain (as discussed above). The supplementary dataset provided includes the measurements of DOC and DBC from all sampling campaigns, the values of each explanatory factor with respect to the upstream catchment of each sampling location and during each sampling campaign, and auxiliary information such as catchment sizes and sampling coordinates (see the acknowledgements section for details regarding access to this data).

Sampling Procedure
Samples of surface water were collected from bridges and boats at each sampling location. The river water was poured through a pre-rinsed funnel into 1.5 litre PET bottles. The bottles were covered and kept in cool, dark conditions. GPS locations were logged using a Garmin GPSMAP 64S handheld device.

Analytical Procedures
The entire methodological procedure, from collection to laboratory analyses, is identical to that used previously by Marques et al. (2017) meaning that samples from all sampling locations included in the current study were processed in the same manner. Here, a summary of the key aspects of the analytical procedure is provided. However, for further information the reader is directed towards the supplementary text S2 (Bao et al., 2017;Brodowski et al., 2005;Coppola et al., 2014;Ding et al., 2014Ding et al., , 2013Ding et al., , 2015Dittmar, 2008;Dittmar et al., 2012aDittmar et al., , 2012b; Dittmar GF/F filters (Whatmann, nominal pore size 0.7 µm), two 150 ml aliquots were subsampled from each sample for DOC analysis, and a further 500 ml of each water sample was acidified to pH2 with HCl (32%, analytical grade). Procedural blanks, including the filtration step, were obtained using ultrapure water. Solid phase extraction (SPE) was then performed on the acidified samples using Bond Elut PPL SPE cartridges (1 g; Agilent Technologies), following a large number of foregoing studies (Dittmar, 2008;Dittmar et al., 2012a;Marques et al., 2017).
All aliquots (for DOC analysis) and solid phase extracted samples (for DBC analysis) were then kept in cool, dark conditions during transport to the laboratory.
In the laboratory, the DOC concentration in the 150 ml aliquots was assessed by automated total organic carbon (TOC) analysis on a Shimadzu TOC 5000 analyzer. Procedural blank samples did not contain detectable quantities of DOC. DBC concentrations were determined following the benzene poly-carboxylic acid (BPCAs) method used in many preceding studies of DBC in aquatic environments (Dittmar, 2008;Dittmar et al., 2012b;Marques et al., 2017;Spencer et al., 2015). This method consists of five steps: first, DOM is eluted from the SPE cartridges using HPLC-grade methanol; second, aliquots of the extract are transferred to glass ampoules and the methanol is evaporated; third, the DOM is re-dissolved in HNO3 and the ampoules are flame-sealed and heated for 9 h at 170•C in a furnace; fourth, the HNO3 is evaporated and the DOM residue is re-dissolved in a phosphate buffer at pH 7.2; finally, BPCAs produced by the partial oxidation of polycondensed aromatic compounds are quantified by ultrahigh performance liquid chromatography (Waters Acquity UPLC) with a photodiode array light-absorbance detector (Dittmar et al., 2012a;Stubbins et al., 2012b). Concentrations of DBC are calculated from the concentrations of benzene penta-carboxylic acid (B5CA) and benzene hexa-carboxylic acid (B6CA) using a well-established power-function relationship (Stubbins et al., 2012b).

Catchment Delineation
The catchment drainage areas upstream of all sampling locations were delineated using the ArcGIS ArcHydro toolbox (Kraemer & Panda, 2009). Void-filled SRTM digital elevation model (DEM) grids with a resolution of ~90 m were used as inputs to the ArcHydro modelling framework, which determines the direction of flow and downstream accumulation of flow paths based on elevation data from the DEM. Streams were defined at a drainage area threshold of 1 km 2 . Sampling locations were mapped to the stream grid generated by ArcHydro and the upstream catchment of each sampling location was computed as areas draining to these locations.

Generalised Linear Modelling
Observed concentrations of DOC and DBC in the river water samples collected in the Brazilian river catchments were modelled using generalised linear models (GLM) fitted to each dataset (GLMDOC and GLMDBC, respectively) using the R statistics package (R Core Team, 2017). Full details of the model fitting procedure are provided in supplementary text S3 (Breheny & Burchett, 2013;Burnham & Anderson, 2002;Fox & Monette, 1992;Fox et al., 2016Fox et al., , 2014Johnson & Omland, 2004;Lakens, 2013;Levine & Hullett, 2002;Mela & Kopalle, 2002;R Core Team, 2017;Sullivan & Feinn, 2012;Zhang, 2016), however a brief outline of the approach to model fitting is provided here. The independent variables included in the model were selected to represent a number of environmental factors which were specified a priori, and thus no stepwise or criterion-based variable selection procedures were followed (Burnham & Anderson, 2002;Johnson & Omland, 2004). The combination of error family and link function of the GLM models was selected so as to maximize explained variance (R 2 ) without compromising the normal distribution of residual errors. GLMDOC was fitted with a Gamma error family and a logarithmic link function, whereas GLMDBC was fitted with a Gaussian error family and a logarithmic link function. Models were validated by assuring that the standard assumptions of residual normality, homoscedasticity, and multicollinearity were satisfied (Fox et al., 2016(Fox et al., , 2014Mela & Kopalle, 2002). Unduly influential outliers, identified by Bonferroni outlier tests for studentised residuals, were iteratively removed (Fox et al., 2014).
Adjusted R 2 values (henceforth R 2 ) were calculated using the 'rsq' package for R following the variance-function-based approach (Zhang, 2016). Partial eta-squared (η 2 p) values were calculated for each independent variable as a measure of effect size. η 2 p measures the (adjusted) effect of a variable as the proportion of explained variance in the dependent remaining after the effects of other factors present in the model have been partialled out. Similar to R 2 , η 2 p is an estimate of the proportion of variance accounted for by an individual effect (Lakens, 2013; Levine & Hullett, 2002;Sullivan & Feinn, 2012). Partial residual plotting was used to illustrate the relationship between each independent variable and the dependent variable given that other independent variables are also included in the model (Breheny & Burchett, 2013). The partial residuals were plotted across the gradients or factor levels of each independent variable whilst assuming that all other continuous independent variables held their median value and that all other categorical independent variables held their modal value.
A GLM model was also fitted to the dataset of DBC/DOC ratios from the samples (GLMDBC/DOC; supplementary material). The same procedure was followed as described above, however a prior step was added in which all outliers excluded from GLMDOC and GLMDBC were also excluded from GLMDBC/DOC. Table 1 provides a summary of the independent variables included in the analysis. For each independent variable, the value assigned to each sample was calculated as the spatiallyaveraged mean within the area upstream of the sampling point; this value was determined using the 'isectpolyrst' function of the Geospatial Modelling Environment (Beyer, 2015).

Land Cover
The MapBiomas collection 2.3 land cover dataset was chosen to represent modern land cover in all catchments (Brazilian Annual Land Use and Land Cover Mapping Project Team, 2017).
The main advantage of using the MapBiomas dataset over other regional mapping options is the consistency of its methodology, classification scheme and resolution (30 m

Soil Organic Carbon, Clay Content and Taxonomy
SoilGrids1km is a global 3D soil model fitted to 110,000 soil profiles, distributed globally,  figure S1). Full details regarding our use of the SoilGrids1km data is provided in supplementary text S4 (Baxter, 2007).

Weather Variables
Precipitation was represented by data from the Tropical Rainfall Measuring Mission (TRMM) multi-satellite precipitation analysis product 3B42RT (Huffman et al., 2010), specifically the accumulated rainfall in the 7 day period prior to the collection of each sample (expressed in mm). Initially, only rainfall occurring in the 2 day period prior to sampling (also from TRMM 3B42RT) was considered as a variable on the basis that all catchments in our dataset had a hydrological lag time of <48 hours (when calculated using catchment slope and catchment length following (Watt & Chow, 1985)). However, rainfall was low in the studied catchments in the 2 day period prior to sampling (data provided in the supplementary dataset): 76% experienced no rainfall; just 8% experienced rainfall in excess of 1 mm and; only 2 catchments (<1%) experienced rainfall in excess of 10 mm. Given these dry conditions immediately prior to sampling, the 7-day rainfall variable predominantly represented the effect of rainfall in the period 3-7 days prior to sample collection. We discuss the implications of this period of influence for our results in sections 6.1 and 6.3.
Surface air temperature was also included on the basis that these broadly correlate with soil temperature at large spatial scales (Smerdon et al., 2006) and affect the rates of SOC and SBC decomposition (Cheng et al., 2008;Sierra et al., 2015). Daily maximum surface air temperature data was extracted from a dataset of gridded meteorological variables in Brazil, which is based upon the interpolation of observational data from 735 weather stations across the country (Xavier et al., 2015). The daily maximum temperatures (°C) were averaged in the 7 days prior to the sampling. The spatial resolution of both the rainfall and temperature datasets was 0.25° (~ 25 km).

Soil Moisture
Soil moisture was represented by satellite observations from the European Space Agency Climate Change Institute (ESA CCI) soil moisture product (Dorigo et al., 2017), which merges active and passive microwave soil moisture retrievals from multiple satellites into a combined product with a spatial resolution of 0.25° and daily temporal resolution (Fang et al., 2016). The daily soil moisture values, expressed in volume of water per volume of soil (m 3 m -3 ) were averaged over the 7-day period preceding the collection of each sample.

Soil Black Carbon
In a meta-analysis of 560 measurements of soil BC and OC concentrations, Reisser et al. (2016) reported a general linear model for BC/SOC ratios that included categorical variables relating to clay content, pH, mean annual precipitation (MAP), mean annual surface air temperature  (Reisser et al., 2016;Willmott & Matsuura, 2000).

Aerosol BC Deposits
Aerosol BC deposition was modelled at the scale of South America using the UK Met Office Hadley Centre Global Environment Model version 2 earth system model (HadGEM2-ES;

K-means Clustering
The  Figure 3 shows the composition of the centroid of each group of catchments that was identified by k-means clustering, in addition to the distribution of the catchments across these groups in each of the ecoregions included in the study. With regards to soil type, the majority of catchments in the Atlantic Forest and Cerrado ecoregions were categorised as cluster A, which was dominated by the Ferralsol soil class. Catchments grouped to cluster D, which was characterised as a combination of predominantly Ferralsol and Nitisol soils, were also present in these ecoregions. A small number of catchments in the Atlantic Forest ecoregion were grouped to cluster B, which exhibited high Cambisol and Ferralsol extent. All Amazonian catchments were grouped into either cluster A or cluster C (Acrisoldominated) soil classes.
With regards to land cover, the catchments were spread across a greater number of clusters in all ecoregions. Cluster C, which was characterised by mixed cover of pasture, forest and some cropland was the most common cover in the Atlantic Forest ecoregion and in Amazonia. A greater number of catchments were grouped to cluster E (highly forested/wooded) in Amazonia than in Atlantic Forest and Cerrado. Cluster D (highly urbanised) was only present in two catchments of the Atlantic Forest ecoregion. Few catchments were assigned to cluster B, which exhibited high cover of cropland, in Amazonia or Atlantic Forest, whereas this cluster was common in the Cerrado. The other cluster with high forest and cropland cover (cluster A), was showed that the SOC stocks, slope, soil moisture, soil classification cluster, clay content and temperature had statistically significant (non-zero) effects on DOC concentration, while land cover cluster had no significant effect (Table 2; Figure 4). GLMDBC also explained a substantial portion of the observed variance in DBC concentrations (R 2 = 0.64) and the RMSE of the predictions was similarly low (0.040 mg L -1 ) in relation to the mean fitted value (0.175 mg L -1 ). SBC stocks, aerosol BC stocks, slope, soil moisture, temperature, soil classification cluster, rainfall and soil clay content all had statistically significant effects on DBC concentration, while land cover cluster had no significant effect (Table 3; Figure 5). The GVIF threshold of 10 (1/(2*df)) was not exceeded by any variable in either GLM, however there were moderate correlations between SOC and clay content, SOC and slope, and soil moisture and rainfall (supplementary figure S2).
GLMDOC and GLMDBC showed a comparable effect of upstream SOC and SBC on the concentrations of DOC and DBC in channels (Figures 4 and 5). In the respective models, increases in upstream SOC and SBC stocks were modelled to produce nonlinear increases in their concentrations. According to η 2 p values, the partial proportion of variance accounted for by the effects of SOC stock and SBC stock in the respective models was 19% and 16% (Table   2 and Table 3, respectively). The similarity of these η 2 p values suggests that DOC and DBC were comparably sensitive to the variability in the upstream densities of SOC and DBC.
The partial proportion of variance accounted for by the aerosol BC term (32%) in GLMDBC was twice greater than that of SBC, which demonstrates that DBC concentrations in these channels were more sensitive to variability in upstream aerosol BC input than they were to upstream stocks of SBC (Table 3; Figure 5). Removing the aerosol BC term from GLMDBC had negative variable was a factor of 1.2 greater in GLMDOC than in GLMDBC, whereas the effect size of soil moisture was slightly lower in GLMDOC than in GLMDBC. In combination with the variables representing stocks of BC and OC, these factors were responsible for the majority of the explained variance in the model, which suggests that these were the primary environmental controls on DOC and DBC export from catchment stocks to channels.
The remaining environmental variables that explained significant portions of the variance in DOC and DBC differed somewhat between GLMDOC and GLMDBC. In GLMDOC, these variables were, in order of reducing η 2 p, soil classification cluster, soil clay content and temperature (Table 2). In GLMDBC these variables were temperature, soil classification cluster, rainfall and soil clay content (Table 3). The main differences in the effect size of the mutually significant variables of the models related to clay content, whose η 2 p was twice greater in GLMDOC than GLMDBC; temperature, whose η 2 p was a factor of 3.5 greater in GLMDBC, and; soil classification cluster, whose η 2 p was a factor of 2.1 greater in GLMDBC. Rainfall was a significant term in GLMDBC but not in GLMDOC. The magnitude of the effects of each independent variable on DOC and DBC concentrations is visible in the partial residual plots shown in figures 4 and 5.
In addition to the greater magnitude of variability in DBC concentration than DOC concentration observed across soil classification clusters, the factor levels over which DOC and DBC varied were not identical in GLMDOC and GLMDBC. The mean DOC concentration was modelled to be significantly higher in channels downstream of catchments assigned to the soil cluster C (Acrisol-dominated), than those grouped into cluster B (Cambisol-dominated).
In contrast, DBC concentrations were modelled to be similar in channels draining catchments in clusters B and C. Moreover, concentrations of DBC (but not DOC) downstream of catchments in both of these clusters were modelled to be lower than those downstream of catchments in clusters A (Ferralsol-dominated) and D (Nitisol-dominated). Overall, the models indicate that the export of DBC from soils was more sensitive to soil mineralogical and physico-chemical properties than DOC.  Table 2: Outputs from GLMDOC, which was fitted to observed DOC concentrations (mg L -1 ). As this model was fitted with a logarithmic link function, it predicts the natural logarithm of DOC concentrations. The mean (± standard error) parameter estimate is shown for each variable. Abbreviations in column headings are as follows: df, degrees of freedom; s. e., standard error; F, the F statistic from an ANOVA type II test for significance of individual model parameters; Pr(>|F|), F statistic significance level; η 2 p, the partial proportion of variance explained by each variable; Z, the Z statistic from a Tukey HSD post-hoc test for significant differences between levels of a factor; Pr(>|Z|), Z statistic significance level. The median, interquartile and extreme values of the distribution of model residuals are also provided (mg L -1 ).

Discussion
This study evaluated the effects of a diverse range of catchment characteristics on the riverine loads of DOC and DBC in a large set of samples from Brazilian rivers. The discussion below first addresses advances made in the understanding of DOC export and then focusses on the apparent coupling of DOC and DBC export processes, drivers of decoupling and the contribution of aerosols to DBC export.

Controls on DOC Concentrations
Overall, 70% of the variability in observed DOC concentrations was explained by the catchment characteristics considered. GLMDOC showed that variability in the concentrations of DOC in the Brazilian river channels was predominantly driven by (i) upstream stocks of SOC and (ii) hydrological factors associated with drainage and the contact time between water and SOC (slope and soil moisture). Small but significant portions of the explained variability were associated with secondary factors linked to (iii) the mineralogy and physico-chemical properties of the soil (soil class cluster and clay content) and (iv) recent temperatures. Neither rainfall nor land cover explained a significant portion of the variability in DOC concentration.
A large number of foregoing studies have shown that upstream stocks of SOC are the dominant source of DOC to freshwater systems (Aitkenhead-Peterson et al., 2003;Dai et al., 2012;Sobek et al., 2007). This dependency is evidenced by the strong relationship that is typically observed between the spatial extent of land covers with rich stocks of SOC, such as peatland and wetland, and the DOC concentrations in channels (Ågren et al., 2007;Dillon & Molot, 1997;Frost et al., 2006;Gergel et al., 1999;Mattsson et al., 2005;Mulholland, 1997;Wilson & Xenopoulos, 2008). Further, direct correlation between SOC stocks and DOC concentrations has been demonstrated at regional to global scales (Hope et al., 1997;Ludwig et al., 1996). The significant effect of catchment SOC stocks on DOC concentrations evident in GLMDOC is thus consistent with foregoing research.
The dominant controls over the rate of DOC export from stocks of SOC were slope and soil moisture, which have also been implicated as important controls on DOC concentration and export fluxes in previous analyses. Due to its controls over soil drainage, flow paths to channels, and the depth of organic soils, slope is an indicator of the contact time between water and SOC stocks (Dillon & Molot, 1997;Ludwig et al., 1996). Increases in catchment slope are associated with enhanced runoff flow generation and shallower organic soils, which promote the rapid delivery of water to channels (Aitkenhead & McDowell, 2000;Dillon & Molot, 1997;Rasmussen et al., 1989). Our work adds to the collection of studies showing an inverse relationship between slope and DOC concentrations in lakes (Rasmussen et al., 1989;Xenopoulos et al., 2003), local to regional -scale river systems (Clair et al., 1994;Frost et al., 2006;Mulholland, 1997), and globally significant catchments (Ludwig et al., 1996). This highlights the role of the hydrological setting, bounded by topography, in modulating rates of DOC mobilisation.
Meanwhile, soil moisture can be considered an indicator of the hydrological state of a catchment on seasonal timescales because it responds in a delayed or smoothed manner to variability in input (precipitation) and output (including discharge and evapotranspiration).
DOC mobilisation depends on soil moisture due to its control on the contact time between drainage water and SOC and thus rates of SOC dissolution. High levels of soil moisture indicate that the water table is high and thus that a greater fraction of the hydrological flow path is in contact with the organic soil horizons where SOC is disproportionately held, enhancing the desorption and dissolution of SOC to DOC (Inamdar et al., 2011;Lambert et al., 2011;Raymond et al., 2016;Singh et al., 2014). This process is compounded by coincident biological responses to soil moisture; specifically, by reducing osmotic stress on microbial cells, high soil moisture also enhances the efficiency with which microbial communities convert substrates to DOC with a relative aversion to re-stabilisation (Davidson & Janssens, 2006;Kaiser & Kalbitz, 2012;Kaiser & Zech, 2000;Malik & Gleixner, 2013;Moyano et al., 2013;Sierra et al., 2015;von Lützow et al., 2006). It was previously shown in headwater catchments that soil moisture anomalies explain the temporal variability in the relationship between SOC stocks and DOC export fluxes (Wilson & Xenopoulos, 2008).
The dominance of SOC stocks, soil moisture and catchment slope as factors explaining variability in DOC concentration helps to explain the strength of the relationship between the spatial extent of wetlands and peatlands and the concentrations or export of DOC that has been observed in a large number of foregoing studies (Ågren et al., 2007;Dillon & Molot, 1997;Frost et al., 2006;Gergel et al., 1999;Harrison et al., 2005;Mattsson et al., 2005;Mulholland, 1997;Wilson & Xenopoulos, 2008). These environments typically store large SOC stocks, occupy regions of low slope, and exhibit poor drainage, thus incorporating the extreme values of the three variables that produce the greatest concentrations of DOC in drainage channels according to GLMDOC.
Among the secondary factors that exerted significant control on DOC concentrations were soil clay content and soil classification cluster. Soil clay content and taxonomy have not previously been included as independent variables in statistical analyses of the variability in riverine DOC concentrations, presumably due to a paucity of data at the required spatial or temporal resolutions. Nonetheless, the potential for soil textural and physico-chemical properties to influence the export of DOC from catchments is well recognised in the literature concerning the cycling of OC in soils. Specifically, OC is cyclically transferred between pools of SOC and DOC by the processes of microbial decomposition, desorption and adsorption (Kaiser & Kalbitz, 2012). This cycle can be halted or slowed via the formation of organo-mineral associations with metal oxides derived from clay weathering or by the physical protection of OC within soil (micro-) aggregates, both of which favour its stabilisation in the SOC pool (Baldock & Skjemstad, 2000;Bruun et al., 2010;Doetterl et al., 2018;Kaiser & Guggenberger, 2000;Kaiser et al., 1996;Kaiser & Kalbitz, 2012;Kaiser & Zech, 2000;Mulholland, 1997;Oren & Chefetz, 2012;Six et al., 2002;von Lützow et al., 2006). Since the retention of OC is, by definition, inversely related to its availability for export, the formation of organo-mineral complexes and micro-aggregates inversely controls the rates of DOC export from catchments.  Inamdar, 2014;Inamdar et al., 2011;Yoon & Raymond, 2012). We found that rainfall was relatively low in the week prior to sampling in the majority of our study catchments, lying within a range of 2-4%, 2-4% and 0-1% of mean annual rainfall in AF1, AMZ1 and the Cerrado sites, respectively (supplementary dataset). This indicates relatively limited potential for rainfall to influence rates of DOC mobilisation in several of our study regions during the period of study. Nonetheless, the range was higher in AF2, spanning 4-14%, 0-0.6%, and 0-8% in the 2013 wet season, 2013 dry season and 2014 wet season, respectively, and it is reasonable to expect DOC concentrations to have shown some variability driven by antecedent rainfall. We suggest that DOC concentrations showed no response to 7-day rainfall because a low fraction of this rainfall occurred within the hydrological lag time (2 days at most; section 4.4.2) prior to the collection of the majority of our samples. While the 7-day rainfall variable predominantly represented rainfall during the 3-7 days prior to sampling, the discharge generated by rainfall during the 3-7 days prior to sampling can be expected to have passed through the sampling locations prior to sample collection.
Overall, the results are highly consistent with previous studies of spatiotemporal variability in DOC concentrations at large spatial scales and outside of extreme precipitation events, which provides confidence that the dynamics of DOC in the rivers studied here are driven by the same processes that have been observed globally in many river systems. While our study focuses on tropical catchments in South America, the consistency of the environmental controls on DOC concentrations identified in our study with the controls identified in other tropical (Huang et al., 2012), temperate (Dillon & Molot, 1997;Frost et al., 2006;Hope et al., 1997;Wilson & Xenopoulos, 2008), high-latitude (Ågren et al., 2008;Clair et al., 1994;Mattsson et al., 2005), and global datasets (Aitkenhead-Peterson et al., 2003;Harrison et al., 2005;Ludwig et al., 1996;Mulholland, 2003) indicates that the results of this study have general relevance to organic matter cycling in global river catchments. Inclusion of soil clay content and mineralogy as explanatory variables also revealed their significant, though secondary, effects on DOC retention and export at large geographic scales.

Coupled Drivers of DBC and DOC Export
This study represents the most comprehensive attempt to model the effects of multiple independent variables on DBC concentrations in river systems. Overall, 64% of the variability in observed DBC concentrations was explained by the independent variables that were considered (R 2 of GLMDBC = 0.64). The primary factors that explained the majority of the variability in DBC concentration were similar to those driving variability in DOC concentrations, specifically: (i) upstream stocks of BC associated with both SBC and aerosol BC deposits, and (ii) hydrological factors associated with drainage and the contact time between water and SOC (slope and soil moisture; table 3; figure 5). Significant portions of the explained variability were also associated with secondary factors linked to (iii) the mineralogy and physico-chemical properties of the soil (soil class cluster and clay content) and (iv) recent weather conditions (temperature and rainfall). Land cover was not modelled to significantly affect the concentration of DBC.
GLMDOC and GLMDBC identify the same primary factors as the dominant drivers of variability in DOC and DBC. Specifically, upstream SBC stocks as well as slope and soil moisture were responsible for the majority of the explained variance in the concentrations of DBC. The consistency between GLMDOC and GLMDBC is a significant result because it suggests that the export of DBC from river channels is dependent on the same set of first-order processes. This result aligns with the hypothesis of Jaffé et al. (2013), who suggested that the processes governing DOC and DBC export must be highly coupled on the basis of the simple linear relationship that exists between their concentrations in a global dataset. Two explanations have previously been proposed to explain the apparent 'coupling' of DBC and DOC export in freshwater systems. Firstly, the solubilisation of SBC to DBC may be a direct result of hydrophobic interactions between SBC and DOC in soil pore space, which lead to DOC acting as an 'agent' for DBC mobilisation. This hypothesis was recently posed and experimentally trialled by Wagner et al. (2017), who suggested that pore-water DOC forms intermolecular associations with SBC through van der Waals forces, hydrogen bonding, or cation bridging.
However, soil leaching experiments did not reveal a relationship between DOC addition to soil and the quantity of DBC leached (Wagner et al., 2017). An alternative explanation for the strong relationship between DBC and DOC concentrations in river systems is that the factors driving the solubilisation of SOC and export of DOC operate synchronously across environmental gradients to solubilise SBC and export DBC from soils. The consistency of the principal drivers of variability in GLMDOC and GLMDBC, and the magnitude of their effects, clearly provides support for the latter hypothesis and moreover suggests that coupling is driven by hydrological factors. Specifically, the synchronised influences of slope and soil moisture on the riverine concentrations of DOC and DBC suggest that the hydrological setting and seasonal-scale hydrological state of these catchment have parallel controls on rates of DOC and DBC mobilisation from soils.

Drivers of De-coupling
Despite the evidence for a general hydrological coupling of DOC and DBC dynamics, there were several discontinuities between GLMDOC and GLMDBC. Differences in the effect sizes of rainfall, temperature, soil clay content and soil classification in GLMDOC and GLMDBC suggest that these environmental factors have differential effects on DOC or DBC export. Further, GLMDBC/DOC revealed that two of these factors (temperature and soil clay content) significantly affected riverine DBC/DOC ratios (supplementary text S7; supplementary table S2). While these environmental factors were all associated with relatively minor portions of the explained variance in DOC and DBC concentration in comparison to stocks, slope and soil moisture, the differential effects of these factors indicate their ability to 'de-couple' the fluxes of DOC and DBC from river systems.
DBC concentrations were more sensitive to recent temperature than DOC concentrations. This observation aligns well with expectations from kinetic theory applied to soil organic matter cycling, which suggests that the temperature sensitivity of decomposition increases with its recalcitrance (Davidson & Janssens, 2006;Dungait et al., 2012;von Lützow & Kögel-Knabner, 2009). As SBC is a particularly recalcitrant form of SOC, its temperature sensitivity is deductively greater than that of bulk SOC, although problems with separating the labile and recalcitrant components of charcoal have impeded the empirical assessment of this phenomenon in incubation studies (Nguyen et al., 2010). A general measure for the temperature sensitivity of SOC decomposition is the proportional change in the rate of SOC mineralisation due to a 10°C temperature increase (Q10 coefficient). The current study provides an indirect indication of the relative Q10 values of SOC and SBC. With all factors other than temperature held at their median or modal value (as in figures 4 and 5), the dQ10 values, defined here as the proportional change in the rate of DOC or DBC export due to a 10°C temperature increase, were 1.5 and 2.3, respectively. If it is assumed that rates of DOC and DBC export from soil each correlate with the rates of their mineralisation, which though speculative is perhaps realistic given the inherent connection of these processes to decomposition, then the ratio between dQ10 values may prove to be a useful predictor of the ratio between the Q10 values that apply to bulk SOC and SBC. The one previous study that measured the temperature sensitivity of BC across an environmental temperature gradient showed a Q10 value of 3.4 (Cheng et al., 2008), which is within the overall range of sensitivities displayed by SOC (~1.5-4;(von Lützow & Kögel-Knabner, 2009)). To our knowledge, no work has yet measured the temperature sensitivity of SOC and SBC simultaneously.
DBC concentrations showed relatively lower sensitivity to soil clay content than DOC concentrations, which contrasts with the hypothesis expressed in previous work that the stabilisation and retention of DBC may be particularly sensitive to clay content (Bruun et al., 2014;Singh et al., 2015;Soucémarianadin et al., 2014). This hypothesis has been held because it is thought that the aromatic structure and polar functionality of BC promotes interaction with clay minerals and uncharged SOC, which in theory make it more prone to protection within organo-mineral aggregates (Czimczik & Masiello, 2007;von Lützow et al., 2006). However, despite these theoretical considerations, contradictory results have been observed in previous research on the effect of clay content on BC cycling in soil. For example, Singh et al. (2015) observed no significant change in charcoal decomposition rate over a clay content range of 8-18%, in contrast to Bruun et al. (2014) who measured a significant reduction across a range of 11-23%. In another study, clay content was found to have a lesser bearing on BC decomposition than soil mineralogy (Fang et al., 2014), which is consistent with the higher effect size of soil classification cluster than clay content in GLMDBC (table 3). It is recognised that the role of clay mineral interactions in the soil dynamics of BC is poorly understood and that more evidence will be required in order to support further interpretation of their effects at large scales (Fang et al., 2018).
DBC concentrations reduced significantly across the rainfall gradient, in contrast to concentrations of DOC which showed no significant trend across the same gradient. Moreover, GLMDBC/DOC suggested that this difference in response had a marginally significant effect on DBC/DOC ratios in the sampled channels, indicating a dilution of DBC fluxes relative to DOC fluxes across a rainfall gradient. We suggest two potential mechanisms for the dilution of DBC relative to DOC as a result of rainfall. First, although the concentrations of SOC and SBC both reduce with soil depth, such that the majority of both stocks are held in the organic layer, studies typically report that SBC is a smaller component of total SOC stocks in the organic horizon than in mineral horizons (Brodowski et al., 2007;Koele et al., 2017;Rodionov et al., 2010;Soucémarianadin et al., 2014). Near-surface flow paths generated by rainfall events might thus be expected to disproportionately mobilise bulk SOC relative to SBC, decoupling the relationship between DBC and DOC in river channels. We do not believe that this mechanism drove the dilution of riverine DBC in our study because the dataset analysed here did not include substantial rainfall events in the 2 days prior to sampling (section 6.1). A more likely explanation for the BC dilution observed in our study relates to the disproportionate effect of antecedent rainfall (3-7 days prior to sampling) on BC and OC mobilisation. SBC stocks are inherently more recalcitrant than SOC (Kuzyakov et al., 2014;Schmidt et al., 2011), which places a lower limit on the rate at which DBC can be mobilised than the rate at which DOC can be mobilised from soils. During a rainfall event, the maximum rate of DBC mobilisation from SBC may fail to keep pace with the rate at which DBC is evacuated from soil pore space, resulting in a relative and absolute depletion of the latent stock of DBC in soil pore space following rainfall. A persistence of this depletion following rainfall might explain why riverine DBC concentrations fell across the rainfall gradient represented in our study.
The 'pulse-shunt' concept refers to the hydrological flushing of DOC from catchments to channels during rainfall or snowmelt events (the 'pulse') followed by the rapid transfer of this DOC through river systems (the 'shunt'; (Raymond et al., 2016)). The related concept of hysteresis refers to variation in the relationship between catchment stocks of organic matter (including SOC) and riverine concentrations of dissolved organic matter (including DOC), which is driven by variability in the composition and quality of catchment organic matter stocks and by the timing with respect to foregoing hydrological events (Evans & Davies, 1998;Vaughan et al., 2017;Wagner et al., 2019). In practical terms, hysteresis means that different fractions of SOC can respond discordantly and on different timescales to a 'pulse'. SBC, as a sub-pool of bulk SOC stocks, has unique molecular properties that limit its rate of decomposition relative to other fractions of SOC and this indicates that, following a 'pulse', the latent stock of DBC in soil pore space is likely to recover over longer timescales than the latent stock of bulk DOC (Hockaday et al., 2007;Kuzyakov et al., 2014). On this basis it is plausible that, for a period following rainfall, pore water DOC is deficient in DBC and the relative stock of DBC available for export to channels is reduced (Bao et al., 2019;Wagner et al., 2015). These concepts may help to explain our observation of a negative relationship between riverine DBC concentrations and antecedent rainfall (which occurred predominantly in the 3-7 days prior to sampling), despite the apparent absence of any rainfall effect on riverine DOC concentrations. Specifically, we suggest that pore water DOC concentrations rebounded following antecedent rainfall, resulting in no observable effect of rainfall on riverine DOC concentrations at the time of sampling, whereas pore water DBC concentrations remained suppressed following the prior rainfall, resulting in a reduction in DBC export from catchments across the rainfall gradient. Overall, our results indicate that seasonal hydrology (represented by soil moisture) and hydrological setting (represented by catchment slope) are critical factors moderating the rate at which SBC stocks are solubilised and made available for export as DBC, while the dilution effect of preceding rainfall on riverine DBC concentrations may result from the flushing of DBC from soil pore space and a deficit in the period that follows. This interpretation is also consistent with the dilution of DBC concentrations observed during highdischarge events in temperate catchments (Bao et al., 2019;Wagner et al., 2015) and with the asymptotic limitation to increases in DBC concentration across seasonal discharge gradients that was previously observed in AF2 (Dittmar et al., 2012a).
Our results yield substantial new insight into the processes driving the export of DBC from catchment stocks of BC in these tropical river catchments, as well as the factors that control the rate of these processes. It remains to be tested explicitly whether a similar array of environmental factors controls the export of DBC from other tropical, temperate or high latitude river systems. Nonetheless, the factors controlling DOC export in the catchments studied here are highly consistent with those operating in diverse environments studied previously (section 6.1) and this suggests that the drivers of organic matter dynamics in these tropical catchments are broadly comparable to those of a wide range of other global catchments.
This, combined with the explicable consistencies and distinctions between GLMDOC and GLMDBC (discussed above), provides a degree of confidence that the environmental factors implicated here are likely to be important controls on the export of DBC in catchments in other global settings. We suggest that these environmental controls should be represented by explanatory variables in future studies of riverine DBC export in other regions of the world

The Influence of Aerosol BC Deposits
GLMDBC showed that the partial proportion of variance in DBC concentration accounted for by aerosol BC deposition outweighed that of SBC stocks by a factor of 2, signifying that DBC concentrations in river channels were more sensitive to upstream stocks of aerosol BC than they were to stocks of SBC. This finding follows a number of recent studies that have demonstrated the potential of aerosol BC to contribute nontrivially towards the DBC load of river channels. The influence of aerosol BC on export has been demonstrated through direct 14 C isotopic methods which identify DBC derived from fossil fuels (Wang et al., 2016), through spatial evidence for DBC fluxes from catchments without considerable charcoal stocks (Ding et al., 2015), and through the physical modelling of DBC inputs to channels (Jones et al., 2017).
In the current study, a statistical model of the factors explaining variability in riverine concentrations of DBC identified aerosol deposition as the independent variable capable of explaining the greatest proportion of variability in DBC concentration in river channels. This implies that BC aerosol deposition has the potential to drive de-coupling of DOC and DBC export, although the stock of aerosol BC deposits was linked to only marginally significant variability in the observed DBC/DOC ratio (table S2).
The strength of the effect of BC aerosol deposition on observed DBC concentrations alone does not suggest that the contribution of BC aerosol to the DBC load of rivers dominates over that of soil BC sources. We previously reported that the contribution of BC aerosol to the riverine DBC load of the Paraíba do Sul River (study region AF1) was most likely to be in the region of 5-18% (Jones et al., 2017), while 15-22% of the riverine DBC load in major Chinese rivers was found to derive from fossil fuel aerosols (Wang et al., 2016). This suggests that the major source of DBC to inland aquatic systems is the soil stock of BC, in agreement with previous studies at the catchment and global scale (Dittmar et al., 2012a;Jaffé et al., 2013). Therefore, the most reasonable explanation for the robust influence of BC aerosol on riverine concentrations of DBC is that soil stocks of BC supply a base flow of DBC to channels, upon which measurable variability is superimposed by the delivery of BC aerosol deposits.
Overall, the rate of aerosol BC deposition to the South American tropical river catchments was observed to explain a greater proportion of the variability in DBC concentration than any other factor included in the analysis, and hence this work supports the developing theory that aerosol deposits make a consequential contribution to fluxes of DBC across the land-ocean continuum.
A recent review of DBC dynamics in aquatic systems suggests that further evidence is required in order to assert that BC aerosol makes nontrivial contributions to riverine DBC at regional scales and worldwide (Wagner et al., 2018). In the current study the effect of BC aerosol was apparent at the continental scale in catchments spanning regions with diverse soil properties, land cover, recent weather conditions and topography.

Active Pipe Processes as Sources of Unexplained Variance
While GLMDOC and GLMDBC explain significant portions of the observed variance in DBC and DOC concentrations, a limitation of our approach was the absence of factors representing inchannel dynamics of DOC and DBC. The concentration of DOC in river channels is not only a function of inputs from upstream catchments, but also of processes occurring in transit such as autochthonous DOC production, exchange of OC between the dissolved and particulate phases, biotic mineralisation and photo-oxidation Stanley et al., 2012;Tank et al., 2010). Some of these processes have also been shown to affect DBC concentrations in some aquatic systems, although at present the wider importance of these processes for BC dynamics in aquatic pools is poorly understood (Coppola et al., 2014;Stubbins et al., 2012b;Wagner et al., 2018). In-channel processing has previously been conceptualised as an 'active pipe' model of a river system, in which a fraction of the riverine carbon is either transferred to river bed sediments or mineralised during transit Raymond et al., 2016;Regnier et al., 2013). This model contrasts with the 'passive pipe' model in which the carbon reaching channels is exported without modification to the global oceans. Our results show that the majority of the observed variance in DBC and DOC concentration is explained by catchment properties, especially hydrological factors, however the unexplained variance of the fitted models might relate in part to active pipe processes that are not represented by our study design.

Conclusion
In this study we investigated the influence of a diverse range of environmental factors on the riverine export of DOC and DBC. We observed that the dominant factors explaining variability in DOC and DBC concentrations are shared. Specifically, soil moisture, slope and the availability of OC and BC, respectively, in upstream catchments were found to be key explanatory variables with comparable effect sizes on each dependent variable. We conclude that the synchronised influences of soil moisture and slope on the export of DOC and DBC from upstream stocks of soil OC and BC, respectively, can explain the coupling of DOC and DBC concentrations in regional and global datasets. These factors relate to the hydrological setting and seasonal state of the catchments and thus indicate that it is principally hydrological factors that drive the coupled dynamics of DBC and DOC in these systems.
Despite this, we also identify a number of environmental factors that weaken the relationship between DOC and DBC. Relative to DOC concentrations, DBC concentrations were significantly more sensitive to temperature and less sensitive to soil clay content. We conclude that spatial and temporal variability in these factors may drive the 'de-coupling' of DBC and DOC dynamics in river catchments. Antecedent rainfall also drove significant variation in DBC concentration while having no significant effect on DOC concentration, which also resulted in marginally significant variation to DBC/DOC ratios. We propose that the delayed recovery of soil pore water DBC stocks relative to pore water DOC stocks following rainfall is a mechanism for the hydrological 'de-coupling' of DBC and DOC export fluxes.
Finally, at an unprecedented geographic scale, our model suggests that aerosol BC contributes significantly towards riverine fluxes of DBC. The fitted models indicated that BC aerosol deposition rates exert a twice-greater effect on DBC concentrations than the magnitude of soil BC stocks, which is most reasonably explained by the superimposition of aerosol-derived DBC fluxes upon a base flow of DBC from soil BC stocks. This result is the strongest evidence yet that aerosol BC deposits make ubiquitous and nontrivial contributions to the riverine load of DBC at large spatial scales and across catchments with diverse hydrology, topography, climate and soil properties.