Volume 120, Issue 7 p. 1274-1287
Research Article
Free Access

The optical properties of river and floodplain waters in the Amazon River Basin: Implications for satellite-based measurements of suspended particulate matter

Jean-Michel Martinez

Corresponding Author

Jean-Michel Martinez

GET, UMR 5563, IRD/CNRS/Université Toulouse 3, Toulouse, France

Instituto de Geociências, Universidade Nacional de Brasília, Campus Universitário Darcy Ribeiro, ICC Centro, Brasília, Brazil

Correspondence to: J.-M. Martinez,

[email protected]

Search for more papers by this author
Raul Espinoza-Villar

Raul Espinoza-Villar

GET, UMR 5563, IRD/CNRS/Université Toulouse 3, Toulouse, France

Instituto de Geociências, Universidade Nacional de Brasília, Campus Universitário Darcy Ribeiro, ICC Centro, Brasília, Brazil

Search for more papers by this author
Elisa Armijos

Elisa Armijos

CLIAMB, Instituto Nacional de Pesquisas da Amazônia-Universidade Federal do Amazonas, Manaus, Brazil

Search for more papers by this author
Luciane Silva Moreira

Luciane Silva Moreira

Departamento de Geoquímica, Universidade Federal Fluminense, Niterói, Brazil

Search for more papers by this author
First published: 18 June 2015
Citations: 37

Abstract

Satellite images can now be used to assess river sediment discharge, and systematic studies over rivers and lakes are required to support such applications and document the variability of inland water optical properties at the watershed scale. The optical properties of the Amazon Basin waters were analyzed from in situ measurements of the remote sensing reflectance (Rrs) at 279 stations and downwelling diffuse attenuation coefficients (Kd) at 133 stations. Measurements of the apparent optical properties, suspended particulate matter (SPM) contents, and characteristics and colored dissolved organic matter (CDOM) absorption spectra were performed during 16 cruises along the main Amazonian Rivers draining the Andes and for some tributaries. Surface-suspended sediment granulometry and mineralogy showed a stable distribution at the catchment scale, even over large distances and between tributaries. The particle number-size distribution was best described using a segmented distribution with a slope of 2.2 for the fine range (1–15 µm), and the CDOM absorption coefficient at 440 nm varied from 1.8 to 7.9 m−1. Overall, both Rrs and Kd were strongly correlated with SPM, although strong CDOM absorption limited the use of the blue spectrum. Reflectance saturation from blue to red was observed at approximately 100 g m−3, whereas the near-infrared (NIR) wavelength enabled the monitoring of the full SPM range (5–620 g m−3). In contrast, Kd showed no saturation for SPM from green to NIR, and a linear model was calculated. The use of the reflectance ratio was investigated and shown to improve the suspended sediment concentration retrieval performance.

Key Points

  • Water optical properties and characteristics were assessed in the Amazon Basin
  • CDOM absorption and SPM show reduced variability (size and type) at river surface
  • AOPs show robust correlation with SPM at infrared for all rivers and tributaries

1 Introduction

Monitoring of inland water quality using remote sensing data represents a major challenge for water-color research due to the complexity of their optical properties relative to oceanic and coastal waters (i.e., Case 1 and 2 waters). In inland waters, absorption and scattering by colored dissolved organic matter (CDOM) and mineral particles can mask the phytoplankton optical properties and show a very weak covariance, thereby invalidating most common retrieval models based on ocean color data that are used over marine water. The optical properties of mineral particles suspended in water depend on the material concentration, size distribution, and refraction index, which, in theory, could demonstrate a site- and/or time-dependent relationship between apparent optical properties (AOPs) and inherent optical properties (IOPs) with the inorganic particulate material.

Suspended sediment fluxes in rivers are the result of erosion, transport, and deposition processes that occur within catchments. The quantification of these fluxes is necessary to monitor and understand the impacts of human activities (e.g., land use and hydraulic infrastructure) and climate change (e.g., extreme events and changes in rainfall patterns) at the catchment scale. However, an accurate calculation of sediment budgets is often difficult due to the poor availability and reliability of sediment flux data in most developed and developing countries [Walling and Fang, 2003]. Water quality-monitoring networks are usually based on water sampling in different locations in a watershed (i.e., hydrological stations) to monitor the production, accumulation, and transfer of certain elements of interest within the river network. The efficiency of such monitoring is therefore a direct function of the number of sampling locations and sampling frequency. In this manner, the capacity of remote sensing imagery to frequently monitor different locations over a short time span may be efficiently used to complete existing field hydrological station networks. Unfortunately, there is a lack of systematic measurements of the main AOPs in continental waters that may support the operational use of remote sensing imagery for water quality monitoring.

A large number of studies have been conducted to measure the relationship between AOPs, IOPs, and parameters of interest (often the pigment concentration), and various syntheses have been published for oceanic waters [Bricaud et al., 1998; Morel et al., 2007; Stramski et al., 2001] and coastal regions [Babin et al., 2003b; Doxaran et al., 2009; Neukermans et al., 2012; Snyder et al., 2008]. These syntheses have paved the way for detailed optical modeling and robust remote sensing monitoring at both the regional and global scales. A significant number of studies have also been published on inland water optical properties [Costa et al., 2013; Giardino et al., 2007; Gitelson et al., 2008; Hoogenboom et al., 1998; Kirk, 1976; Ma et al., 2006; Whitlock et al., 1981]. However, unlike the progress that has been achieved for marine waters, there is a lack of systematic studies of rivers and lakes documenting the variability of the AOPs/IOPs in these areas to establish the foundation for the remote sensing-based operational monitoring of inland waters. In particular, it is necessary to consider the whole watershed, instead of a specific river/lake, and to determine the variation of AOPs/IOPs as a function of the hydrological cycle.

In recent years, several studies have shown that medium-resolution remote sensing imagery (few hundred meters per pixel) such as Moderate Resolution Imaging Spectroradiometer (MODIS) or Medium-Resolution Imaging Spectrometer (MERIS) may be efficiently used for the monitoring of the suspended sediment in large rivers [Heege et al., 2014; Martinez et al., 2009; Park and Latrubesse, 2014; Wang and Lu, 2010] and lakes [Kaba et al., 2014; Majozi et al., 2014; Odermatt et al., 2010; Wu et al., 2013]. Martinez et al. [2009] showed that robust empirical relationships between the suspended particulate matter (SPM) and surface reflectance using corresponding MODIS 250 m images and field samples collected over 7 years can be derived with significant accuracy (30%) and an absence of seasonal bias at an hydrological station along the Amazon River in Brazil. Using 10 years of MODIS data for the Amazon River in Peru, Espinoza Villar et al. [2012] compared upstream and downstream remote sensing-derived river sediment discharge estimates and demonstrated that satellite assessments are robust (root-mean-square error of 18%). Mangiarotti et al. [2013] analyzed methods of combining conventional network data and MODIS-derived SPM estimates to improve sediment budget assessments in the Amazonian Plain. However, all of these studies were based on empirical relationships between remote sensing reflectance (Rrs) and SPM, which may prevent a generalized use of reported remote sensing methods. In the context of a dramatic decrease of sediment monitoring stations across the world [Covault et al., 2013], the use of remote sensing may be an efficient alternative to follow the fate of sediment discharge in river catchments. Therefore, it is necessary to investigate the robustness of the link between the optical properties of water and suspended sediment to ensure the robustness of the remote sensing-based retrieval.

The objective of this study is to understand how the optical properties of surface water vary across a large river basin and to determine the significance of this variability based on a large data set of water sampling and optical measurements. In particular, we assessed the variability of the remote sensing reflectance and of the vertical attenuation coefficient for different periods of the hydrological cycle, over floodplain lakes and river mainstreams as well as between different subcatchments, focusing on the mineral fraction of the SPM. This study also contributes to the field of bio-optic modeling by extensively documenting the optical properties of the world's largest watershed.

2 Materials and Methods

2.1 The Amazon Basin

The Amazon catchment covers 5.9 × 106 km2 [Callède et al., 2010] encompassing various soils and lithologies. The Amazon Basin is composed of three major morphostructural units: (i) the Brazilian and Guyana cratonic shields, (ii) the Andes and Subandean zones to the west, and (iii) the lowlands where the main rivers progressively join to form the Amazon River. The cratonic shields cover approximately 44% of the Amazonian Basin area and are composed of strongly weathered and metamorphosed igneous rocks. The Andes accounts for 11% of the Amazonian Basin area and consists of Paleozoic to Meso-Cenozoic metamorphic, igneous, and sedimentary rocks. The Amazonian lowlands cover approximately 45% of the Amazonian Basin and are composed of a very thick layer of Cenozoic sediments.

Interestingly, the first and most common classification used to describe the main Amazonian water types is based on their coarse optical properties, catalogued as white, clear, and black [Sioli, 1950]. White waters refer to waters loaded with sediment showing strong dissolved organic material, mainly originating from the Andes. Clear waters originate from the cratonic shields and show very low SPM loads but also low levels of organic matter. Black waters are of low biological productivity with near-zero SPM concentrations, have almost no nutrients, and are strongly acidic due to the organic material present in the colloidal fraction, which also gives them their dark color. The main black-water rivers drain the northern tributaries and cratonic shields, such as the Negro River and the Trombetas River, although black waters can also be found in a large number of smaller subcatchments in the Amazon lowlands that finally converge in the Amazon River main stem from the piedmont to the Atlantic Ocean. The white waters flow through the Madeira River and Solimões catchments that drain the southern Andes and the central and northern Andes, respectively. Herein, we mainly focus on these sediment-dominated waters along the Madeira River, the Solimões River, and the reach of the Amazon River that is formed by the Madeira, Solimões, and Negro Rivers.

The Amazon Basin crosses two hemispheres and experiences contrasting rainfall regimes. Consequently, the hydrological regime shows a distinct high-water period from March to April in the Madeira catchment in the south and from June to July in the Negro catchment in the north. The Amazon/Solimoes River shows an intermediate hydrological cycle as it receives waters from both the north and the south of the catchment [Richey et al., 1989]. Overall, the hydrological regime becomes more consistent from upstream to downstream. From the Andes piedmont to the lowlands, the flood regime is monomodal with a steadily increasing water level during the rainy season. Consequently, during the majority of the hydrological cycle, the suspended sediment and geochemical characteristics of the surface water change slowly. The HYBAM network data (www.ore-hybam.org) allow for assessments of the suspended sediment cycle throughout the catchment. Variations in SPM concentration are correlated with the river discharge in the Andes [Armijos et al., 2013], whereas the SPM becomes progressively disconnected with the river discharge through the confluence with other rivers draining the forested catchments that provide almost no SPM to the system [Filizola and Guyot, 2009]. Consequently, there is an overall decrease of the SPM concentration from upstream to downstream due to sedimentation and dilution processes [Meade et al., 1979]. Several studies have noted that the dissolved organic carbon (DOC) concentrations show a low seasonal difference in the main stem of the river [Ertel et al., 1986; Hedges et al., 1986; Moreira-Turcq et al., 2003] of between 3 and 5 g m−3, whereas in black waters, the DOC concentrations are typically greater than 10 g m−3. The proportion of humic acids versus fulvic acids, known to affect the extent of light absorption, has been shown to remain relatively constant seasonally, with a ratio of fulvic to humic acids in the Amazon River at Óbidos of 3.2 ± 0.3 [Ertel et al., 1986]. The Amazon Basin is marked by a large floodplain that significantly contributes to the water fluxes and mass transfer. The open floodplain lakes are affected by different processes, such as the mixing of different water masses derived from the local catchments and the rivers, resuspension processes during the low-water period, and intense blooms of phytoplankton during the dry season. Such phytoplankton blooms extend over large areas during the period of decreasing water levels from July to September. The phytoplankton composition is dominated by Cyanophyceae with a significant proportion of Bacillariophyceae and Chlorophyceae [Nogueira et al., 2010].

2.2 Calculation of Apparent Optical Properties: Rrs and Kd

Figure 1 shows the locations of the 42 stations where radiometric measurements were made from 2007 to 2011. Below- (133 measurements) and above-water (279 measurements) radiometric measurements were performed using TriOs RAMSES radiometers operating in the 350–900 nm spectral range. One radiometer was mounted with a cosine collector for irradiance measurements, and two other radiometers were equipped for radiance measurements with a field of view of 7° in air. All of the radiometers were synchronized to simultaneously measure the various optical properties.

Details are in the caption following the image
Location of the 42 stations for above-water (279 measurements) and below-water (133 measurements) sampling in the Amazon River Basin from 2007 to 2011 along the two main rivers draining the Andes: the Solimões and Madeira Rivers. Stations are partitioned into four subgroups: (1) the Amazon River and its main tributary, the Solimões River; (2) the Madeira River; (3) other tributaries; and (4) floodplain lakes.
For the majority of the locations, radiometric measurements were collected both from the boat deck to calculate the remote sensing reflectance Rrs (λ) and in the water to retrieve downward irradiance Ed(z,λ) profiles. Above-water measurements of the upwelling radiance Lu (λ) were preferred to below-water measurements Lw(λ) due to the strong light absorption in the blue and near-infrared (NIR) domains by water, which may result in unreliable reflectance assessments at those wavelengths. The remote sensing reflectance was calculated as Rrs (λ):
urn:x-wiley:21699003:media:jgrf20414:jgrf20414-math-0001(1)
where Ed(λ) is the downwelling irradiance above the water surface, Lu(λ) is the upwelling radiance above the surface water, and Ls(λ) is the sky radiance that is used to correct for the skylight reflection effect at the air-water interface. The above-water upwelling radiance Lu is the sum of the upwelling radiance Lw(0+) and the sky radiance directly reflected by the air-water interface Lr. Because only Lu is directly measureable, and Lw(0+) and Lr are not measured, Lr is assessed as Lr = ρ.Ls, where ρ is a proportionality factor. The factor ρ is not an inherent optical property of the surface and is dependent on the sky conditions, wind speed, solar zenith angle, and viewing geometry. Mobley [1999] used a radiative transfer code to estimate the variability of ρ as a function of the different forcing factors. These results showed that when Lu is acquired with a viewing direction of 40° from the nadir and 135° from the Sun, the variability of ρ is considerably reduced under clear-sky conditions, and a value of 0.028 is acceptable at wind speeds less than 5 m s−1. Finally, Lw(0+) was assessed by the subtraction of Lu and ρ.Ls. To limit the effects of external factors, all radiometric measurements were acquired within the viewing geometry defined by Mobley and under low-wind conditions (0–4 m s−1) and clear-sky conditions and for Sun zenith angle values ranging from 0 to 30°.

The in-water downward irradiance Ed(z,λ) and upward radiance Lu(z,λ) profiles were measured at discrete depths in rapid succession within the euphotic layer, which was usually less than 3 m in depth. The diffuse attenuation coefficients for downward irradiance Kd(λ) were calculated by comparing the depth with the slope of the natural log plot for the downwelling irradiance. To retrieve robust estimates of Kd at each station, at least two vertical profiles were collected sequentially, and the more stable profile was selected.

2.3 Water Sampling

In situ water quality data were collected during 16 cruises along the main Amazonian Rivers in Peru and Brazil from 2007 to 2011 in different seasons (see Table 1). This sampling scheme made it possible to register the radiometric variability of the waters in the Amazonian catchment, the river main stems and floodplains and the estuary, representing locations separated by more than 3000 km. Although the sampling program was focused on the two rivers draining the Andes (Amazon and Madeira Rivers), samples were acquired from the tributaries and floodplain connected to these two streams to document the radiometric variability among the black and white waters. When possible, the above and in-water measurements were realized sequentially within approximately 15 min, with the surface water samples collected during each measurement for water quality assessments.

Table 1. Location and Dates of the 16 Sampling Campaigns Conducted for This Study
Campaign Time Period Water Stage Sampling Location Number of Stations Visited
Above Water In Water
1 30/01/07 20/02/07 Rising Santarem to Macapá 2 1
2 21/05/07 25/05/07 High Manacapuru 19 0
3 10/11/07 15/11/07 Low/Rising Manacapuru to Borba 34 2
4 15/03/08 20/03/08 Rising/High Manacapuru to Borba 58 2
5 10/04/08 10/04/08 Rising Amazon estuary 2 0
6 15/05/08 23/05/08 High Manacapuru to Santarem 9 8
7 05/10/08 13/10/08 Low Manacapuru to Santarem 14 5
8 20/06/09 06/07/09 High Manacapuru to Santarem 18 13
9 19/11/09 28/11/09 Rising Porto Velho to Manacapuru 7 11
10 26/03/10 30/04/10 High Manacapuru to Porto Velho 5 7
11 08/06/10 02/07/10 Recession/High Iquitos to Santarem 18 18
12 25/08/10 12/09/10 Recession Manacapuru to Santarem 31 14
13 26/01/11 02/02/11 Rising Manacapuru to Santarem 4 5
14 15/02/11 25/02/11 Rising Manacapuru to Santarem 11 8
15 07/07/11 30/07/11 Recession Porto Velho to Santarem 25 20
16 23/11/11 18/12/11 Low/Rising Iquitos to Santarem 22 19
Total 279 133

At each of the visited stations, a single water sample used to determine SPM concentrations was collected at the surface from a small boat or research vessel during each Rrs and Kd measurement. It is important to note that for most of the locations, the water depth exceeded 20 m and reached up to 60 m at some stations. Water samples were processed on board the ship immediately after collection. When using small ships, water samples were stored in 2 L polyethylene containers and processed on land no later than 12 h after sampling. The samples were filtered using 0.45 µm cellulose acetate filters (Millipore) that were previously dried for 24 h at 60°C and weighted. After filtration, the filters were dried for 24 h at 60°C and weighed again to determine the concentration of suspended matter. Phytoplankton pigment concentrations were measured during 4 cruises by filtering 250 mL of water through GF/F filters, which were then frozen at −20°C on board the research vessel. At the laboratory, the pigments were extracted in acetone and analyzed for Chl a and phaeopigment concentrations using a calibrated fluorometer. Herein, Chl is defined as the sum of Chl a and phaeopigments.

Particle-size distribution analyses were performed on 39 5 L water samples collected during 6 cruises. Granulometric data were obtained using a laser grain-size measurement device (Mastersizer 2000 with a sample dispersion unit) at the Companhia de Pesquisa de Recursos Minerais (CPRM) Laboratory for Sediment and Water Quality. The results were modeled according to the classical Junge hyperbolic law to assess the variability of the slope coefficient and the range of sizes over which the fit could be performed.

Samples were collected during 3 cruises to assess the variability of the CDOM absorption coefficient for different types of waters and along an upstream to downstream profile of the Solimões and Amazon Rivers. Water was filtered under low vacuum on a 0.45 µm Millipore membrane using an all-glass filtering device. The filtered samples were stored and frozen in dark 10 cL bottles. At the Universidade Federal Fluminense (UFF) laboratory, a 5 cm quartz cuvette was used to measure the absorbance of the filtered water between 350 and 750 nm at 1 nm increments using a dual-beam spectrophotometer (Shimadzu UV/VIS UV-1800). Milli-Q water was used as a reference. After conversion to absorption coefficients, an exponential function was fitted between 350 and 500 nm by nonlinear regression to the data to assess the exponential slope SCDOM. A baseline correction was not applied to obtain a null CDOM absorption at 685 nm as is usually performed for marine waters [Babin et al., 2003b].

We analyzed 33 samples collected during 8 cruises to determine the temporal and spatial variations of the surface-suspended sediment mineral assemblage along the Madeira and Amazon Rivers. Samples were collected at the river surface in 25 L canisters and evaporated to a final volume of 500 mL. At the laboratory, water samples were prepared by evaporating the aqueous suspensions to dryness on a glass slide. The suspensions were prepared by the ultrasonic dispersion of 10 mg of sample in 2 mL of deionized water. The identification of minerals was based on their reactions to commonly used treatments: air drying, ethyl-glycol solvation, and heating to 550°C. Based on the glycolated sample diffractograms, semiquantitative estimates of the different clay minerals were performed manually by measuring the characteristic reflection areas.

3 Results

3.1 Variability of the SPM Characteristics: Concentration, Granulometry, and Mineralogy

The SPM values of the water samples spanned more than 2 orders of magnitude (2–621.6 g m−3), with a mean value of 140.8 g m−3. There was a general decrease of the SPM from upstream to downstream areas along the rivers draining the Andes (the Maranon, Ucayali, Napo, Solimões, and Madeira Rivers). For each location, the highest SPM values were observed during the rising water period, and a decrease occurred during the remainder of the hydrological cycle. Black-water rivers showed very low SPM levels (i.e., <10 g m−3), whereas in floodplains waters, SPM concentrations showed low to intermediate SPM values (0–76 g m−3). Studies have reported that white-water rivers show low particulate organic carbon (POC) contents (1–4 wt % relative to the SPM concentration) [Moreira-Turcq et al., 2003]. These relationships have been confirmed for a wide range of riverine environments throughout the world, demonstrating that the POC/SPM ratio decreases with increasing SPM, reaching approximately 1% for SPM levels of greater than 100 g m−3 [Coynel et al., 2005; Meybeck, 1993]. In the Amazon River, it has been shown that the POC concentrations directly vary with the discharge volume, and higher concentrations occur during high-water periods (corresponding to low SPM concentrations). In black and floodplain waters, the POC concentrations are much higher, with weight percentages of up to 60%.

The nonalgal particle (NAP) scattering properties (i.e., the absolute values and spectral dependence) vary with the particle-size distribution and may influence the AOP/IOP relationship with the SPM [Babin et al., 2003a]. Suspended sediment granulometry was obtained to assess the variability of the particle-size across the different white-water rivers, at different locations along the main stems and during different seasons. To characterize the particle-size distribution, hydrologists often use the D50 parameter, called the median grain size, which is the grain diameter at which half of the sample is smaller and half is larger. Studies of hydrologic optics use the number-size distribution N(D), which represents the number of particles for a given size interval around the mean diameter D. It is often assumed that N(D) follows the Junge distribution as a function of increasing particle size with
urn:x-wiley:21699003:media:jgrf20414:jgrf20414-math-0002(2)
where K determines the scale and j is the slope of the distribution. Typical values of the factor “j” reported for marine waters (including organic and mineral fractions) vary between 2 and 5 [Mobley and Mobley, 1994]; although most studies assume a unique value for j of 4, experimental data are very limited. The value of this coefficient has been shown to be determinant because when j < −4, the relative contribution of the small particles to the total particle volume concentration decreases, and therefore, the contribution of small particles to the total light scattering decreases [Babin et al., 2003a]. The composition of the Amazon white waters is, however, very different that than of the open ocean and coastal waters; thus, we tested the variability of the slope factor among the different rivers studied herein. Mobley and Mobley [1994] and Morel and Ahn [1991] suggest that the particle-size distribution in water is best described by a segmented distribution with a smaller value of j for smallest particles and a larger value of j for the largest particles. Our data demonstrate the same pattern with a slope of 2.22 ± 0.19 in the 1.1–15 µm range and 4.56 ± 0.61 at greater than 15 µm (see Figure 2). The limited variation in the size distribution for all of the studied samples is supported by the low D50 parameter variability of 12.0 ± 2.84 µm. Interestingly, no seasonal or spatial dependencies have been detected among the collected water samples.
Details are in the caption following the image
Particle-size distribution measurements for 39 surface water samples collected along the Madeira, Solimões, and Amazon Rivers. Segmented power law functions were fitted for the small (<15 µm) and large particle (>15 µm) ranges.

We analyzed 33 samples collected during the field campaigns to determine the temporal and spatial variations of the surface-suspended sediment mineral assemblages. Our results confirmed those of previous reports [Guyot et al., 2007] and showed a balanced composition of quartz (10–20%), kaolinite (20%), illite (20–40%), and smectite (20–40%). We detected a limited but significant increase in the smectite composition from low water to high water, which is consistent with a stronger contribution of lowland tributaries and floodplains to the main stemflow during the high flood phase [Konhauser et al., 1994; Moreira et al., 2013]. Comparisons between the Madeira and Solimoes systems showed higher smectite content in the Solimoes system and a higher illite content in the Madeira River, whereas the quartz and kaolinite remained stable. These observations highlight a limited but significant variability in the type of suspended sediment that may be related to the seasonal and/or spatial dependency of the reflectance over time.

3.2 Light Absorption by Dissolved Organic Matter

The water absorption is conventionally modeled as the sum of the absorption by pure water, SPM, phytoplankton, and CDOM. Unlike Case 1 and coastal waters, continental waters are marked by strong CDOM concentration originating from the degradation of terrestrial organic matter. Figure S1 in the supporting information shows the aCDOM variation for 19 samples collected from black, clear, and white waters as well as floodplain lakes. The aCDOM was higher in the black river waters (aCDOM(440) = 7.9 m−1, N = 1) and floodplain lakes (aCDOM(440) = 4.8 ± 0.8 m−1, N = 2) and lower in white (aCDOM(440) = 3.3 ± 1.0 m−1, N = 14) and clear waters (aCDOM(440) = 1.8 ± 0.8 m−1, N = 2). Interestingly, at longer wavelengths, the CDOM absorption in white river waters remained significant (aCDOM (670) = 0.72 ± 0.5 m−1). The spectral slope SCDOM showed small variations (0.013 nm−1 ± 0.0016) from one location to another. The CDOM absorption classes followed the usual water-type classification used to describe the main water types in the Amazon (clear/white/black), for which both DOC and POC have been shown to exhibit increasing carbon concentrations. The aCDOM values reported here were much stronger than in any other measurements for the open ocean and can only be compared with coastal waters marked by large freshwater inputs from land. Bowers et al. [2000] monitored a CDOM absorption gradient from the Clyde Sea to an estuary and measured an aCDOM value of 1.6 at zero salinity. The fact that SCDOM does not show a strong variation is in accordance with other studies in which this parameter was assessed in other coastal waters. Babin et al. [2003b] reported a mean SCDOM value of 0.018 ± 0.002 over 345 samples collected from six different coastal regions, and Bowers et al. [2000] reported a mean value of 0.018 ± 0.005 for 25 samples. In a previous work, Bricaud et al. [Bricaud et al., 1981] found an SCDOM value of 0.014 ± 0.003, and Kirk [1976], who studied lake waters, observed SCDOM values between 0.013 and 0.016. In an analysis of CDOM in a tidal marsh area, Tzortziou et al. [2008] showed that SCDOM was systematically lower at low tide (0.0149), as dissolved terrestrial material was drained out of the marshes, than during high tide (SCDOM = 0.0168). Our study did not allow to document systematic variations of aCDOM across the watershed, although our observations could be used to quantify the strong CDOM attenuation in the Amazonian waters and show the relative stability of those values for white-water rivers.

3.3 Classification of the Rrs Spectra

A cluster analysis was performed to classify the entire set of Rrs spectra into homogeneous groups. The objective of this classification was to highlight spatial or temporal dependencies in the remote sensing reflectance data set that may support the use of multiple retrieval algorithms for different rivers or a global retrieval algorithm. For this analysis, we used the k-means unsupervised classification technique, which categorizes the data set into classes based on the natural distribution of the data in multivariate space. In contrast to supervised classification techniques, this approach does not require predefined classes. To define the optimal number of classes, we used a hierarchical cluster method by applying Ward's algorithm to squared Euclidian distances between the spectra [Lubac and Loisel, 2007]. To enhance the spectral shape of the Rrs spectra in the classification, each Rrs spectrum was previously normalized by its integral calculated over the entire spectrum. The initial data set of 279 spectra was reduced to 171 spectra to avoid the overrepresentation of the Solimões River and Madeira River samples. The analysis of the hierarchical clustering showed that the partition into eight classes minimized the within-cluster variances. Figure S2 shows the mean spectrum of each class, and Figure S3 shows the mean and standard deviation for each class. Each group is described below.

Class 1 consisted of the Rio Negro black waters (N = 3). This class presented the lowest reflectance values of <0.003 sr−1 (nonnormalized reflectance values), with a slightly increasing reflectance from blue to red wavelengths. This pattern was directly related to the high CDOM concentrations associated with low Chl and SPM values. Class 2 included black- and clear-water stations primarily from the floodplain (N = 4). The class spectrum slowly increased from 400 to 570 nm and then declined gradually toward the higher wavelengths. Chl a absorption at approximately 665 nm was clearly apparent, although the other characteristic Chl a absorption peak at 440 nm was not detected. This class presented the lowest SPM loads and high Chl concentrations (mean = 34.4 mg m−3). Class 3 mainly consisted of the productive floodplain waters (N = 21) and samples from river confluences. For this class, the reflectance increased sharply from 400 to 575 nm and presented a wide maximum up to 700 nm. The local minimum observed at approximately 675 nm most likely corresponded to the presence of Chl a. A second reflectance maximum was detected at 800 nm, and this class presented intermediate SPM values (mean = 21.9 g m−3) and high Chl concentrations (mean = 21.5 mg m−3). Class 4 was mostly composed of floodplain waters and river confluence stations (N = 28) and showed similar spectral patterns and reflectance values to those of Class 3, although the first reflectance maximum was shifted toward larger wavelengths at 585 nm, and the reflectance did not exhibit marked absorption troughs up to 700 nm. Class 4 was associated with intermediate SPM values (mean = 23.4 g m−3) and lower Chl a concentrations (mean = 9.5 mg m−3) compared with Class 3.

Class 5 consisted of floodplain waters (N = 6) and presented two distinct maxima of comparable amplitude: one near 550 nm and one near 705 nm. The first peak matched the Chl a absorption minimum in the green spectrum, whereas the second peak corresponded to the Chl a absorption minimum in the red spectrum. Two local minima detected at 625 and 675 nm corresponded to the absorption maxima of phaeopigments and chlorophyll a pigments, respectively. An absence of pigment-induced absorption patterns was observed in the blue part of the spectrum, which was likely because of the strong absorption by CDOM. Not surprisingly, Class 5 showed the highest Chl concentrations (66.9 mg m−3) among all classes and an intermediate SPM value (mean = 39 g m−3).

Classes 6–8 included river sediment-dominated waters with strong SPM concentrations. All of these classes showed similar spectral patterns, with sharply increasing Rrs values from 400 to 580 nm and a much less pronounced increase up to 700 nm. Class 6 (N = 57) presented intermediate SPM values (72 mg m−3) and a low Chl value of 1.8 mg m−3. Class 7 (N = 33) showed strong SPM values (187 mg m−3), and Class 8 (N = 19) showed the highest SPM values (368 mg m−3), which was associated with low Chl a concentrations (<1 mg m−3). It is important to note that Classes 7 and 8 can only be discriminated based on the shapes of their NIR spectra, although they showed different SPM values. Interestingly, although the spectra were normalized to enhance the spectral shape in the classification, the Rrs classes represented homogeneous classes as a function of SPM or Chl, and spatial or temporal dependencies were not detected. These conclusions support the use of a global SPM retrieval algorithm for the entire catchment, which we will be detailed in the next two sections.

3.4 The Relationship of Rrs and Kd to SPM: Average Relationship

Figure 3 shows the variation of the remote sensing reflectance at 670 nm and 850 nm as a function of the SPM for 279 samples collected during the 16 cruises in of all the studied rivers and floodplain lakes. Although the reflectance in the red and NIR channels showed a clear increasing trend as a function of the SPM, we noted a strong difference between the channels. Although the red channel showed saturation at approximately 100 g m−3, we could not identify a saturation point for the NIR channel up to 600 g m−3, demonstrating a remarkable sensitivity to the SPM concentration over 2 orders of magnitude. The dispersion around the main trend appeared to be stronger for lower SPM values, specifically for the floodplain stations. Linear regressions between the natural logarithm of the wavelength and of Rrs for the Madeira River and Solimões River showed similar intercepts (P = 0.763) and slopes (P = 0.461). In contrast, the slopes and intercepts for Madeira River and Solimões River were different when compared with the tributary and floodplain samples (P < 0.001). However, the lower SPM range for the floodplains (mean SPM = 22.9 g m−3) compared with that for the Madeira (mean SPM = 239.3 g m−3) and Solimões (mean SPM = 128.9 g m−3) Rivers limits the relevance of this statistical result. The clear correlation between SPM and reflectance for all of the water types and seasons demonstrated that SPM is clearly driving the water optical properties in the majority of the waters of the Amazon Basin. However, whereas the SPM was mainly composed of mineral particles at high concentrations and along the white-water rivers, the composition of the suspended matter may be much more diverse at low SPM values with a much larger organic fraction (living and nonliving), in particular in black river and floodplain waters. Relative importance and variability of the organic fraction relative to the inorganic fraction clearly affected the reflectance level at low SPM concentrations (i.e., <20 g m−3).

Details are in the caption following the image
Variation of (top) Rrs(670) and (bottom) Rrs(850) as a function of the SPM concentration. The collected samples are identified based on their sampling location: the Solimões and Amazon Rivers (triangles), the Madeira River (black circles), the floodplain lakes (white circles), and all of the tributaries (squares).

Figure 4 shows the variation of Kd(650) as a function of the SPM for 133 samples collected during the 16 cruises in all of the studied rivers and floodplain lakes. Unlike the Rrs at the red wavelength, saturation was not observed up to 620 g m−3. The dispersion was relatively constant across the entire SPM range, and a simple linear function could be fit to model the Kd-SPM relationship. At low SPM levels, Kd(650) appeared to be greater than 1, which is indicative of residual CDOM absorption at those wavelengths (see section 3.2). This result showed that Kd is a good predictor of SPM for all types of waters in the Amazon Basin.

Details are in the caption following the image
Variation of Kd(670) as a function of the SPM concentration.

We also investigated the spectral dependence of the relationship of Rrs and Kd with SPM, excluding all floodplain and nonwhite waters. The revised data set represented 229 stations for Rrs and 109 stations for Kd. To assess the accuracy of the AOPs in predicting SPM concentrations, two subsets were formed for each AOP. For Rrs, the calibration data set included 115 samples (mean SPM = 168.2 g m−3) and the validation data set included 114 samples (mean SPM = 166.2 g m−3). For Kd, the calibration data set included 56 samples (mean SPM = 128.1 g m−3) and the validation data set included 53 samples (mean SPM = 138.9 g m−3).

Figure 5 presents the model-fitting and SPM retrieval performances for Rrs and Kd as a function of the wavelength. For Rrs, there was a clear increasing relationship with SPM from the blue to NIR wavelengths, with a square Pearson correlation factor of greater than 0.75 beyond 730 nm and an optimal range of 840–900 nm (r2 of approximately 0.81). A lower correlation in the visible spectrum (400 to 700 nm) was induced by the rapid saturation of the Rrs as a function of the SPM at low concentrations. The least squares regression showed that a power law model a.Rrs(λ)b provides the best fit values. The model factors and performance were calculated based on the linear regression between the natural logarithm of the wavelength and of Rrs. The root-mean-square error (RMSE) of the retrieval model decreased as a function of increasing wavelength starting at 114.1 g m−3 (400 nm) down to 61 g m−3 (820–900 nm). The optimal wavelength occurred at 860 nm, and the corresponding model is as follows:
urn:x-wiley:21699003:media:jgrf20414:jgrf20414-math-0003(3)
Details are in the caption following the image
Variation of the square of the Pearson correlation coefficient (left axis) and the RMSE of the SPM concentration (right axis) as a function of the wavelength for Kd and Rrs.
Figure S4 shows the model performance over the validation data set. For Kd, there was a strong and robust correlation with SPM (see Figure 5) over a wide range of wavelengths in both the visible and NIR spectra from 495 to 860 nm (r2 > 0.90). The lower Pearson correlation at short wavelengths and beyond 890 nm may have been induced by the strong absorption of light that makes it difficult to accurately assess the downwelling irradiance distant from the water surface. Accordingly, a linear best fit model between SPM and Kd performed better than for Rrs with an RMSE of better than 30 g m−3 from 500 to 870 nm.
urn:x-wiley:21699003:media:jgrf20414:jgrf20414-math-0004(4)
Figure S5 shows the model performance for the validation data set. These observations may be related to the link between the AOPs and IOPs (the absorption and scattering coefficients); i.e., whereas Rrs is related to the absorption to scattering ratio, Kd is assumed to vary as the sum of absorption and scattering. With increasing SPM, coefficients a and b will increase, meaning that Kd will increase indefinitely as a function of the suspended sediment concentration, whereas Rrs may reach a saturation point. Not surprisingly, Kd and Rrs were well correlated at 730 nm up to 900 nm (r2 > 0.80). At 850 nm, we observed the following relationship (r2 > 0.85):
urn:x-wiley:21699003:media:jgrf20414:jgrf20414-math-0005(5)

3.5 Improvement of SPM Retrieval Using the Rrs Ratio at Two Wavelengths

It has been shown that the ratio between Rrs at two wavelengths may lessen the dependence on the SPM type and thus is more robustly linked to the SPM concentration [Doxaran et al., 2002; Moore et al., 1999]. Irradiance reflectance is generally modeled as R = f · bb/(a + bb), where bb is the backscattering coefficient and f is a factor determined by the incident light field and often assumed to be invariant as a function of the wavelength. Doxaran et al. [2002] analyzed the variation of the bb/(a + bb) ratio for highly turbid waters in the Gironde estuary. These authors showed that when the contributions of pure water, algal pigments, and CDOM to light absorption can be neglected, the bb/(a + bb) ratio becomes independent of the SPM concentrations, leading to the saturation phenomenon observed for Rrs. The total absorption by pure water is high at NIR, and thus, this simple model predicts that saturation will occur for higher SPM concentrations at infrared relatively to visible wavelengths. Interestingly, the same authors reported that the bb/(a + bb) ratio at two wavelengths will be less sensitive to the refraction index and grain-size variation than to the SPM concentration at high SPM loads. Indeed, by assuming that the spectral variations of bb are weak [Moore et al., 1999; Whitlock et al., 1981], it can be shown that the ratio is much more dependent on the absorption than on the backscattering, which, in turn, may limit the influence of the refraction indices and granulometry. This simple ratio model will perform better using wavelengths at which SPM absorption shows a strong variation and, together with the pure water absorption, dominates the total absorption. It can be deduced from our previous observations that the Rrs ratio should make use of the red and NIR wavelengths.

We investigated whether the use of the reflectance ratio Rrs(λ1)/Rrs(λ2) improved the SPM-Rrs relationship in the river sediment-dominated waters. To determine the optimal position of λ1, we selected an initial value of λ2 = 670 nm within the range of the minimum total absorption in the red spectrum. We performed a regression of the model [Rrs(λ1)/Rrs(670)] versus SPM for the range of 730 to 900 nm; the RMSE of the SPM estimation showed minimal values over a wide range of λ1 between 760 and 820 nm (see Figure S6a). In the second iteration, we determined λ2 after fixing λ1 at 805 nm and regressing the model [Rrs(805)/Rrs(λ2)] versus SPM. Again, the RMSE was minimal over a wide range of λ2 between 650 and 700 nm, with the best result obtained at λ2 = 680 (see Figure S6b). Interestingly, the ratio model performed significantly better than with the use of a unique wavelength with an RMSE of approximately 38 g m−3 (Figures S7 and S8). In addition, the spectral bands showing a minimal RMSE at both the red and NIR regions of the spectrum included the spectral bands of commonly used water-color sensors, such as MERIS and MODIS, and future platforms, such as Sentinel-3, NPP/VIIRS, and JPSS.

We further investigated whether the efficiency of the Rrs ratio was global or varied from one river to another. For the Solimões River, the use of the ratio compared with that of a single-wavelength model at NIR increased slightly the model accuracy (r2 = 0.83 versus 0.73, N = 76). For the Madeira River, the use of the ratio showed a strong improvement of the retrieval accuracy (r2 = 0.93 versus 0.62, N = 102) (see Figure S9). However, when the regression lines assessed for both rivers were compared, the data supported the hypothesis that the slope (P = 0.108) and intercept (P = 0.225) are equal. From our data set, we concluded that the use of the Rrs ratio may not show an increased SPM retrieval efficiency globally, although it may be used as a general tool to homogenize a data set across a watershed.

4 Discusion and Conclusions

The Amazon Basin presents Case 2 waters with an absorption budget dominated by CDOM and SPM in rivers but with a significant contribution from Chl a in floodplains. A simple unsupervised classification of Rrs spectra made it possible to discriminate several classes: (1) one class presenting very low Rrs values corresponding to the so-called black waters with strong CDOM absorption and very low SPM concentrations; (2) four groups of floodplain lakes and river confluences, where a reduced water velocity allows the development of phytoplankton at high concentrations; and (3) three classes of sediment-dominated waters with relatively high SPM concentrations. Rrs shows saturation, with SPM at approximately 100 g m−3 in the green and red spectra. However, the use of Rrs in the infrared spectrum makes it possible to monitor the SPM range without saturation. Interestingly, the three main Amazonian Rivers having sediment-dominated water (the Madeira, Solimões, and Amazon Rivers) showed a similar behavior as a function of SPM, even though the samples were collected at stations that in some cases were separated by more than 3000 km. Kd proved to be a very robust predictor (r2 > 0.9) of SPM through a simple linear relationship for a very wide range of SPM values (up to 622 g m−3) and wavelengths.

The use of reflectance ratios has been previously presented as a means to reduce the sensitivity of Rrs to the sediment type or size distribution in sediment-dominated waters [Doxaran et al., 2002; Moore et al., 1999]. In our data set, we observed that the use of a ratio model does not improve the results uniformly for all of the studied water bodies. For example, although the retrieval model accuracy was slightly improved for the Solimões River when using a reflectance ratio instead of one wavelength, for the Madeira River, the improvement is much stronger. In an analysis of 11 years of MODIS red and infrared surface reflectance data and of SPM data acquired at two hydrological stations along the Madeira River in Brazil, Espinoza Villar et al. [2013] showed that the NIR reflectance does not exhibit a simple relationship with SPM. The authors detected a hysteresis effect in which for the same SPM concentration, the reflectance during rising water is higher than at the flood peak. It is important to note that using the same method, Espinoza Villar et al. [2012] did not detect such a seasonal variation for the Solimões River or its main tributaries (Marañon and Ucayali) in Peru.

For such sediment-dominated waters, it is expected that the AOP/IOP variations may be strongly impacted by spatial and seasonal changes in sediment type and particle-size distribution. However, our results showed that at the Amazon catchment scale, Rrs exhibits a relationship with SPM that is sufficiently robust to be used for the accurate monitoring of surface-suspended sediment discharge in the largest river of the world. Regarding the particle-size distribution, our granulometric results showed a relatively consistent grain-size distribution among different rivers and hydrological periods. The segmented grain-size distribution showed a more even distribution (slope 2.22) for the 1 to 15 µm diameter range than is usually predicted in optical modeling (approximately 4), which implies that the contribution of very fine particles (<1 µm) to NAP scattering is limited. These observations agree with the classical results of sediment transport dynamics in a turbulent flow system that predict a relatively constant distribution of the fine-grained size fraction in the water column in which the coarse-grained size fraction increases from the surface to the river bottom [Rouse, 1938], resulting in an almost pure fine sediment composition at the river surface. Furthermore, most large rivers, such as the main rivers of the Amazon Basin, have been shown to display reduced grain-size range in the bed load and suspended sediment mainly because they flow across vast plains for a major part of their course with very low topographic gradient causing strong sedimentation in the foothills and floodplains [Latrubesse, 2008].

Our study did not allow for a detailed assessment of the variability of the mineral particle types, but it did confirm the results of previous studies in the Amazonian Plain showing a relatively consistent pattern of mineral compositions. Guyot et al. [2007] analyzed 229 sediment samples from throughout the entire Amazon Basin to document the spatial variability of the mineral assemblages. These authors reported that the sediment clay phase from the rivers draining the Andes contains a dominant illite + chlorite assemblage that is enriched in smectite during transport through the lowlands. The main minerals found in the SPM showed a limited refraction index range relative to water of 1.16–1.18. Furthermore, previous modeling studies [Babin et al., 2003a; Doxaran et al., 2009; Wozniak and Stramski, 2004] showed that even grain-size distribution (i.e., slope below 4) limited the impact of the variability of the refraction index. Our results demonstrated that the sediment-size distribution and mineralogy may present limited variations at the scale of a large river basin, thus supporting the development of robust SPM retrieval methods from main AOPs.

Although the Rrs seasonal variation as a function of SPM for the Madeira River case can be efficiently corrected using a simple reflectance ratio model, the cause of the Rrs variability remains unclear because the deviation may originate from variations of the physical properties of the inorganic SPM or variation in the organic fraction. Viers et al. [2008] analyzed the variations in the geochemical and physical properties of SPM from the Solimões River and Madeira River during one hydrological cycle. These authors showed that both rivers exhibit distinct suspended sediment chemical compositions but that these characteristics do not vary significantly over time except for in the Madeira River, where the strontium (Sr) isotopic composition showed clear variations between the high-water period and the remainder of the year. Santos et al. [2014] published a more exhaustive monitoring of the dissolved Sr isotopes for the main Amazonian Rivers based on the HYBAM water database. These authors confirm that only the Madeira River catchment waters exhibited a seasonal dependence in the dissolved strontium isotopic composition that can be tracked from the Andean piedmont downstream to the Amazon River confluence. This seasonal variability is attributed to increasing rates of erosion of Precambrian rocks during the rainy season in the upstream catchments (Beni River). This seasonal control of the Sr isotopic composition highlights the limited but significant variability in suspended sediment types in the Madeira River. This variability may be used to explain the seasonal dependency of the reflectance in this subcatchment and the lower Rrs values observed during high-water periods in particular.

This work is one of the first studies to assess the variation of the AOP of river waters across a large river basin and for all hydrological periods. The SPM retrieval performance assessed based on the AOPs varied from 37% (Rrs at a single wavelength in the NIR) to 23% (Rrs ratio at two wavelengths) to 18% (Kd in the visible spectrum) relative RMSE. These performances are consistent with the operational monitoring of suspended sediment in rivers. Uncertainties regarding both the sampling and calculation methods are known to affect the accuracy of in situ SPM concentration assessments in rivers by approximately 10 to 20% [Cheviron et al., 2011; Horowitz, 2003]. In recent years, we have published several studies that used MODIS satellite imagery to retrieve surface SPM values for the Amazon Basin using correlations between Rrs and SPM data obtained from field samples. Martinez et al. [2009] found a 36% relative difference when comparing 7 years of MODIS-derived surface SPM estimates and 10 day surface field samples acquired at one station in Brazil. At two stations in Peru (Solimões and Marañon Rivers), Espinoza Villar et al. [2012] reported a 27% and 45% relative RMSE when comparing 4 years of MODIS-derived surface SPM estimates and 10 day surface field samples. The retrieval accuracy assessed based on the satellite data is consistent with the Rrs-based model performance we analyzed and demonstrates that significant improvement may be achieved using an appropriate reflectance model. The optical properties presented in this study cannot be extrapolated for global use over all the continental waters. However, considering the size of the catchment, our results support the notion that remote sensing monitoring may be used efficiently for sediment discharge monitoring in large river basins.

Acknowledgments

This work has been supported by the Brazilian Water Agency (ANA), CNPq, Ministry of Science and Technologies, and Program for Graduate Students-Agreement (PEC-PG) program in Brazil; IRD, Joint International Laboratory “Environmental Changes Observatory in Amazon Region,” INSU “Reliefs,” INSU “Syster,” ANR “CARBAMA,” and CNES/TOSCA programs in France. Our thanks also go to all the HYBAM program partners in Brazil and Peru: the University of Brasilia, the Amazonas State Federal University (UFAM), the Fluminense State University (UFF), the Brazilian Mineral Resources Research Company (CPRM), and the Peruvian Hydrologic and Meteorology Service (SENAMHI). The authors are grateful to the ship crews who accompanied them during the surveys. The data for this paper are available by contacting the corresponding author.