Spatial linkages between coral proxies of terrestrial runoff across a large embayment in Madagascar

Coral cores provide vital climate reconstructions for site-specific temporal variability in river flow and sed­ iment load. Yet, their ability to record spatial differences across multiple catchments is relatively unknown. Here, we investigate spatial linkages between four coral proxies of ter­ restrial runoff and their relationships between sites. Coral cores were drilled in and around Antongil Bay, the largest bay in Madagascar, and individually analysed for fifteen years of continuous luminescence (G /B), B a/C a, 51 8 Osw and S13C data. Each coral core was drilled close to individ­ ual river mouths (< 7 km), and proxy data were compared to modelled river discharge and sediment runoff data for the three corresponding catchments. A reasonable agreement be­ tween terrestrial runoff proxies with modelled river discharge and sediment yield was observed. Some inconsistencies be­ tween proxy and modelled data are likely linked to proxy behaviour, watershed size and local environmental physiochemical parameters. In general, the further a coral resided from its river source, the weaker the proxy relationship was with modelled data and other corals, due to mixing gradi­ ents and currents. Nevertheless, we demonstrate that two coral B a/C a and luminescence (G /B) records influenced by the same watershed are reproducible. Furthermore, a strong B a/C a relationship was observed between two cores from distant watersheds, with baseline averages in agreement with modelled sediment runoff data. As humic acids behave con­ servatively in the water column, luminescence (G /B) data gave the highest regional correlations between cores, and showed the most consistent relationship with site specific modelled discharge. No statistical relationship was observed between cores in terms of interannual 51 8 Osw and S1 3 C, meaning corals were recording a localised signal at their re­ spective sites, confounded by vital effects. Comparing proxy baseline averages and mean seasonal cycles provided a good overview of the runoff dynamics of the bay system.


Introduction
Anthropogenic and climate-induced changes in sediment load entering the coastal realm are of great concern for the sustainability of tropical marine and terrestrial environments (Rogers, 1990: McClanahan and Obura, 1997: McCulloch et al., 2003: Payet and Obura, 2004).Deforestation often leaves soils susceptible to erosion (Green andSussman, 1990: Agarwal et al., 2005)  of both sediment and leached dissolved components deliv ered to the coastal ocean (Warrick and Rubin, 2007).M ada gascar is an iconic example of the extreme environmental impacts human deforestation and habitat destruction has on soil runoff and land degradation (Green and Sussman, 1990;Harper et al., 2007).It is now estimated that only 10-15% of the original forests remain since extensive deforestation began in the mid 20th century (Green and Sussman, 1990;Harper et al., 2007).Forest protection and management can help stabilise soils within catchment areas, yet requires con tinuous records of site specific land-use changes, erosion and weather patterns to differentiate between vulnerable and sta ble areas.
In Madagascar, weather station data are scarce (Dewar and Wallis, 1999;Dewar and Richard, 2007) and satellite de rived rainfall data around coastal regions with high cloud cover are often unreliable (Quartly et al., 2007).Previous research efforts in Madagascar have focussed on terrestrial environments, yet an assessment of the status of the coastal marine ecosystems in relation to climatic and anthropogenic stressors is lacking (Goodman and Benstead, 2003;Kremen, 2003).Proxy climate and environmental records preserved in annually-banded massive corals, such as Porites spp., can significantly augment the instrumental data that are often too short to identify change in many tropical regions.M as sive corals can grow for centuries at relatively fast rates (1 -2 cm y r-1 ), incorporating trace elements (TE) into the arago nite skeleton according to their relative concentrations in am bient sea water during calcification (Alibert et al., 2003;M c culloch et al., 2003;Lewis et al., 2007;Jupiter et al., 2008;Jones et al., 2009).Such properties make massive corals ideal archives of localised environmental change (e.g.river discharge, sediment load).The down-core analyses of coral proxies in long coral cores can provide information on sitespecific temporal variability in river flow and sediment loads influencing corals.Such information can potentially assist in the management of watersheds in Madagascar where instru mental data on water characteristics are lacking.However, the reliability of coral proxies is still debated as cores from the same region can often show varying signals (Jones et al., 2009;Pfeiffer et al., 2009;Lewis et al., 2011).
Luminescence of the coral skeleton is used as a tracer of temporal variability in river flow.First described by Isdale (1984), the intensity of luminescent lines in corals was thought to be caused by the skeletal incorporation of humic acids (HA) derived from hinterland soils.Subsequent reports indicated that luminescence may also result from changing skeletal densities (Barnes andTaylor, 2001, 2005).More re cently, spectral luminescence scanning (SLS) has shown that both processes contribute to luminescence and that humic acids can be normalised for the effects of changing skeletal density to provide an indicator of humic acid runoff (Grove et al., 2010).As the luminescent emission signal of HA is slightly longer than aragonite, taking the green/blue (G /B) ratio gives an estimate of the amount of HA relative to the skeletal density (Grove et al., 2010).SLS resolves density effects associated with luminescence intensities, such as de clining trends in intensity with coral age (Lough, 2011a, b;Jones et al., 2009).
Barium (Ba) is both dissolved in the rivers and adsorbed to suspended sediments (clay minerals), which are then trans ported to coastal waters via rivers.As salinities increase, Ba desorbs from the suspended sediment due to the higher ionic strength of seawater.As Ba is diluted by seawater it is thought to follow a conservative mixing pattern (Sinclair and McCulloch, 2004), and thus Ba acts as a tracer for riverine sediment input to the coastal ocean.As Ba substitutes Ca in the coral skeleton, sediment discharge is reconstructed ac cordingly using skeletal B a /C a ratios (Sinclair and McCul loch, 2004;Alibert et al., 2003;McCulloch et al., 2003;Fleitmann et al., 2007).However, as estuarine processes, such as phytoplankton uptake and resuspension, can lead to a non conservative behaviour of Ba (Hanor and Chan, 1977;Coffey et al., 1997); subsequently, sediment discharge reconstruc tions can be affected (Sinclair, 2005).In such circumstances, skeletal Ba / Ca levels may not be directly related to sediment discharge.
Coral skeletal S180 is a function of both SST and salinity.Calculating the difference between coralline 5180 and the sea surface temperature proxy S r/C a provides a salinity proxy, 51 8 Oseawater (51 8 Osw) (Sinclair and McCulloch, 2004: M c Culloch et al., 1994, 2003).Indeed, here we couple S180 and S r/C a following the method of Ren et al. (2002) to recon struct the 518 O of seawater (51 8 Osw).Potentially, 51 8 Osw can identify the coral experiencing the lowest and highest salini ties.This in turn can enhance our interpretation of the other runoff proxies by factoring in mixing processes.
The S r/C a ratio of the coral aragonite seems to be the most robust paleo-thermometer, whereby a negative relation ship exists with SST, i.e. as temperatures increase, less Sr is incorporated into the aragonite lattice relative to Ca (Alibert andMcCulloch, 1997: DeLong et al., 2007).However, re cently published coral S r/C a records covering the past hun dreds of years indicate specific problems with the S r/C a thermometer, particularly on decadal to secular time scales (e.g.Linsley et al., 2004, 2006: Quinn et al. 2006).Pfeiffer et al. (2009) showed that the intrinsic variance of the single core S r/C a time series differs from core to core, limiting their use for absolute estimates of past temperature varia tions.This inter-colony variability seems linked to vital ef fects.
The DIC of riverine waters is typically more isotopically depleted than seawater S1 3 C, caused by the decay of strongly depleted terrestrial vegetation (Moyer, 2008;von Fischer and Tieszen, 1995;Moyer and Grottoli, 2011;Marin-Spiotta et al., 2008).The input of riverine DIC to the coastal ocean will therefore cause depletions in the S13C of ambient seawater DIC and coral skeletons.A reduction in incident light levels may also play a role in determining the skeletal 513 C vari ability (Grottoli, 2002;Grottoli and Wellington, 1999).As sediment and humic acid concentrations increase with runoff, turbidity reduces the incident light reaching benthic commu nities, including the corals (Tarsen and Web, 2009).Dur ing photosynthesis the endosymbiotic algae (Symbiodinium sp.) preferentially utilise 12C for biomass production, leav ing the carbon pool, used by the coral for calcification, en riched in 13C (Weil et al., 1981;Swart et al., 1996;Reynaud-Vaganay et al., 2001;Reynaud et al., 2002).Reduced photo synthesis will reduce the depletion of 12C in the carbon pool and thus the skeletal material will have a lighter S1 3 C, yield ing an inverse relationship with increasing runoff.
In this study, we collected four coral cores from Porites colonies across three watersheds surrounding Antongil Bay in eastern Madagascar.We first examine the reproducibility of the common river runoff proxies B a /C a and Luminescence (G /B ) for two corals associated with the largest w a tershed.Secondly, we examine the relationships of four coral proxies indicating river flow, sediment load, salinity and turbidity/DIC (luminescence (G /B ), B a/C a, á 1 8 Osw, S1 3 C) for coral cores located in the same region, yet associated with three separate river systems.We test whether individual prox ies reflect a regional river runoff signal, a localised signal or a combination of both.As the temporal variation of river dis charge recorded by corals is a function of the distance from a river source (salinity gradient) and the flow direction, as well as the source input, comparing rivers using corals from different reefs is challenging (King et al., 2001;Tough et al., 2002;Carricart-Ganivet et al., 2007;Jupiter et al., 2008;Prouty et al., 2010).To test how representative proxies are of their corresponding river watersheds, we combine coral proxies and compare results with modelled river discharge and sediment yield.

Research area and climate setting
Antongil Bay is located in the NE of M adagascar (Fig. 1) covering an area of 2800 km2, with a mean depth of 41.5m and a coastline of 270 km extending 80 km inland (Ersts and Rosenbaum, 2003).Almost all populated areas are located in the northern and western coastal regions of the bay, in cluding the largest urban areas of Maroantsetra and Mana- nara.Two protected forest areas are found in the vicinity of the bay, the Makira Forest and the Masoala Peninsula Na tional Park.The Makira Forest extends over 4600 km2 north of Maroansetra, and together with the Masoala Peninsula Na tional Park forms one of the largest continuous rain forest areas remaining in Madagascar (Birkinshaw and Randrianjanahaiy, 2007).Since the introduction of the National Park there has been a significant reduction in the rate of deforesta tion, yet it still remains a constant threat to the marine and terrestrial environments (Harper et al., 2007).
One of the largest rivers draining into Antongil Bay is the Antainambalana, running through the Makira For est Area from a source 1450 m above sea level (Good man and Ganzhorn, 2004).Its watershed covers an esti mated 4000 km2 and the river mouth is located in the city of Maroantsetra (Fig. 1 and Table 1).The coral cores MASI and MAS3 were collected 40 m apart next to the island Nosy Mangabe, ca.7 km from the river mouth, on the edge of a fringing reef slope and reef flat, respectively, at approxi mately 4 m water depth (Fig. 1 and Table 1).The ANDRA coral core was collected 30 km from M ASI from the fring ing reef slope on the east side of the bay, ca.7 km from the Ambanizana river mouth, at 4 m water depth (Fig. 1 and Table 1).The Ambanizana has a much smaller watershed  (Collins, 2006;Kremen et al., 1999).The IFAHO coral is situated ca.4.5 km from the mouth of the river Anaovandran (watershed: ca.180 km2) outside of Antongil Bay (Fig. 1 and Table 1).The river origi nates from the same mountains as the Ambanizana, but flows eastwards away from Antongil Bay.For the last ca.11 km it runs through a plain before entering the ocean (Windley et al., 1994 and references therein).The Anaovandran drains a hinterland that consists of granitic soils at high elevations and basaltic rocks that cover the lower elevation partly defor ested sedimentary plain (Douglas, 1967;Collins, 2006;Kre men et al., 1999).The IFAHO coral core was drilled at 6 m water depth and collected from the back reef of a fringing reef where direct influence of the open ocean is restricted.
The climate in M adagascar can be divided into a August-December cold-dry season and a January-Jtrly warm-wet season (Fig. 2).Air temperatures peak in December and January and are lowest between July and September (Kre men, 2003).The sea surface temperature (SST) peaks be tween December and April, reaching maximum average temperatures of 28.7°C (Fig. 2).Minimum average SST are 24.6 °C, meaning there is a mean seasonal range of 4.1 °C.Highest rainfall occurs between January and March, while lowest rainfall occurs between September and Novem ber (Xie and Arkin, 1996).Average rainfall levels reach 300 mm month-1 during the wet season (January to March) and 50 mm month-1 in the dry season.Average peak river discharge occurs between January and April.River discharge then decreases yet continues all year round, reaching lows in October and November (Fig. 2).

Coral sampling
Coral cores were drilled from massive colonies of Porites spp. at depths between 4-6 m at different locations in and around Antongil Bay, NE Madagascar (Fig. 1 and Table 1) during M arch and April 2007.A commercially available pneumatic drill (Rodcraft 4500) was used to extract 4 cm diameter cores along the central growth axis of the colony.In contrast to the results of Matson (2011), drill holes were not plugged as algal growth was observed on the concrete plugs from earlier drilling campaigns.Drill holes closed by natural coral growth within 2yr when left.Cores were sec tioned lengthwise into 7 mm thick slabs, rinsed several times with demineralised water, blown with compressed air to re move any surficial particles and dried for more than 24 h in a laminar flow hood.Annual density bands were visu alised by X-radiograph-positive prints, and the growth axis of the coral slab was defined as the line perpendicular to these bands.The average growth rate of all four coral cores was approximately 1 3 ± 2 m m y -1 (Grove et al., 2010 and Ta ble 1).With a diamond coated drill, subsamples were taken every 1 mm parallel to the growth axis, equivalent to ap proximately monthly resolution.All coral slabs were cleaned with sodium hypochlorite (NaOCl, 10-13% reactive chlo ride: Sigma-Aldrich Company) for 24 h, after sub-sampling for geochemistry, to remove residual organics that might quench luminescence (Nagtegaal et al., 2012).

Coral luminescence
SLS was performed, as described by Grove et al. (2010), on bleached coral slabs using a line-scan camera fitted with a Dichroic beam splitter prism, separating light emission in tensities into three spectral ranges; blue (B), green (G) and red (R).The four coral cores used for this study were also analysed in the study of Grove et al. (2010) to test the per formance of the SLS technique.Digital core images were analysed with the Line Scan Software Version 1.6 (Avaatech), providing RGB intensity values for individual pixels of 71pm in length.Since the spectral emissions of humic acids are slightly longer than aragonite, spectral G /B ratios reflect the changing humic acid/aragonite ratios within the coral skeleton (Grove et al., 2010).The G /B datasets for the four corals applied in this study all begin in January 1991 and end in December 2005, spanning a total of 15yr.

Coral geochemistry
Two measurement techniques, solution ICP-MS and LA-ICP-MS, were applied to compare B a /C a data generated for two coral cores, M ASI and MAS3 (LA-ICP-MS only), from the same catchment.Both time series begin in January 1991 and end in December 2005, spanning a total of 15 yr.Solu tion ICP-MS was also applied to generate both B a /C a and S r/C a data for the cores M A SI, ANDRA and IFAHO for the same 15yr period, at an approximate monthly resolution.Subsamples were taken every 1 mm equating to ca. 1 month resolution.This is in contrast to the approximate sub-weekly resolution of the LA-ICP-MS technique.Stable isotope data were generated for the same three corals measured using so lution ICP-MS (MASI, ANDRA and IFAHO), at the same ca.monthly resolution (subsample resolution).

Solution ICP-MS
S r/C a and B a /C a were analysed for M A SI, ANDRA and IFAHO by high resolution inductively coupled plasma mass spectrometry (HR-ICP-MS; Thermo Scientific Element-2) equipped with a double spray chamber and Teflon mi croflow nebulizer.The method of sample dissolution, di lution, quantification, and drift correction is described in Nagtegaal et al. (2012).Accuracies were determined us ing a JCp-1 Porites sp.coral standard.Accuracies for S r/C a were 99.8 ± 0.5 % (8.80 mmol mol-1 ) and for B a/C a 9 7 ± 6 % (7.16pm olm ol-1 ).Short-term (< 5 min) preci sion and longer-term precision ( 8 batches measured during 1 month) were typically 0.2 % and 0.5 % RSD for S r/C a , re spectfully, and 0.2% and 6 % for B a/C a. Short-term preci sion reflects variability in operating conditions such as power and gas flow rates.Longer-term stability mainly reflects the goodness of drift corrections.Blanks were kept low using ultrapure acids.Blanks were always < 0.5 % for S r/C a and < 2 % for Ba / Ca.

LA-ICP-MS
Laser ablation inductively coupled mass spectrometry (LA-ICP-MS) analytical methods were identical to those reported in Jupiter et al. (2008).The coral slabs of cores M ASI and MAS3 were mounted on a stage containing standards and analysed using a Varian 820 inductively coupled mass spec trometer.The laser slit size was 40 x 400 pm, yielding a sam pling resolution of 20 nm.The resultant data were normalised to 4 3 Ca using a Varian laser scanning analysis software pro gram developed at the Australian National University (ANU) Research School of Earth Sciences (by L. Kinsley).Data were first smoothed using a 1 0 point running mean to reduce the influence of outliers, followed by a 1 0 point mean to re duce data volume.To determine accuracy, a NIST 614 glass standard and a pressed coral standard were used (Fallon et al., 1999(Fallon et al., , 2002)).The overall analytical precision for B a /C a was 4.3% (Fallon et al., 1999).Daily and long-term (5 month) reproducibility was monitored by repeated measurements of the pressed coral standard and an in-house coralline sponge standard (Fallon et al., 1999).The daily and long-term repro ducibility was 1.6% and 3.3% , respectively.The analytical internal precision for B a/C a was < 4.3 % RSD (Fallon et al., 1999).Further details on the methodology and standards are available in Fallon et al. (1999Fallon et al. ( , 2002)).The B a/C a (LA-ICP-MS) datasets for the two corals M ASI and MAS3 begin in January 1991 and end in December 2005, spanning a total of 15yr.

Stable isotopes
To measure the skeletal 5180 and 51 3 C, approximately 80 pg of coral powder was reacted with H3 PO4 in an automated carbonate reaction device (Kiel IV) connected to a Finnigan MAT 253 mass spectrometer at the Royal Netherlands Insti tute for Sea Research (NIOZ).For the NBS19 carbonate stan dard (standard values are -2.2%o and 1.95 %o for S180 and S1 3 C, respectively) we obtained values of -2.21±0.08%o and 1.96 ± 0.04 %o for S180 and S1 3 C, respectively, relative to the Vienna PeeDee Belemnite (V-PDB) standard.Solution ICP-MS B a / Ca, S r/C a and the stable isotope datasets for the three corals M A S I, ANDRA and IFAHO all begin in January 1991 and end in December 2005, spanning a total of 15yr.

Reconstructing salinity
To calculate 51 8 Oseawater (51 8 Osw), we followed the method of Ren et al. (2002).The method assumes that coral S r/C a is solely a function of SST and that coral 5180 is a func tion of both SST and 51 8 Osw.This method separates the effects of 51 8 Osw from SST by breaking the instantaneous changes of coral 5180 into separate contributions by instan taneous SST and 5180 changes, respectively.We used the 51 8 Ocorai-SST relationship of -0.2%o°C (Juilliet-Leclerc and Schmidt, 2001) and the Sr/C a-S ST relationship of -0.06 mmol mol-1 °C (Corrège, 2006) to calculate the in stantaneous changes in SST.Taking into account the uncer tainties associated with the analytical precision of (1 ) coral S1 8 0 , (2) coral S r/C a, (3) the coral S r/C a-S ST calibra tion and (4) the coral 5180-S ST calibration, we calculated a monthly 51 8 Osw error of 0.09 %o (following Zinke et al., 2008).Based on the local salinity-51 8 Osw relationship for Antongil bay, the error in reconstructed salinity equates to 0.67 psu for monthly values and 0.19 psu for mean annual values (Sect.2.7).

Hydrological modelling
Due to a lack of hydrological data for M adagascar we used model results of river discharge and sediment yield in the Antongil Bay (Maina et al., 2012).Maina et al. (2012) used the STREAM (Spatial Tools for River basins, Environment and Analysis of M anagement options) grid-based hydrolog ical model to simulate monthly river discharge (Aerts and Bouwer, 2002;Aerts et al., 1999).STREAM simulates the water balance for each grid (50 km resolution) using a lim ited number of parameters, including spatial-temporal pre cipitation and temperature trends, elevation, land-cover and soil water storage capacity specific to Madagascar (Aerts and Bouwer, 2002).The STREAM model has been suc cessfully applied in various climate and hydrological stud ies (e.g.Bouwer et al., 2006;Winsemius et al., 2006), which have shown that a monthly time step is sufficient for detect ing decadal, inter-annual and seasonal hydrological cycles.
Here, we extracted the average seasonal cycle in river dis charge for the three rivers studied and the average yearly river discharge per catchment (Maina et al., 2012).Sediment yield per unit area was computed using the Non-Point Source Pollution and Erosion Comparison Tool (N-SPECT) developed by the National Oceanic and Atmo spheric Administration (NOAA) (http://www.csc.noaa.gov/digitalcoast/tools/nspect/index.html).N-SPECT combines data on elevation, slope, soils, precipitation, and land cover to derive estimates of runoff, erosion, and pollutant sources (nitrogen, phosphorus, and suspended solids) from across the landscape, as well as estimates of sediment and pollutant ac cumulation in stream and river networks.Erosion rates and sediment loads were calculated using the Revised Universal Soil Loss Equation (RUSLE) and Modified Universal Soil Loss Equation (MUSLE) (Wischmeier and Smith, 1978).N-SPECT provides annual mean (January-December) esti mates of sediment yield per unit area in m ty r-1 at a 50 km resolution.Here, we extracted the average yearly river sedi ment yield per catchment (Maina et al., 2012).

Coral age model construction
Coral chronologies were based on the seasonal cycle of B a /C a and G /B ratios.Luminescence in corals has pre viously been used to cross date core records (Hendy et al., 2003;Grove et al., 2010) and assist in establishing a chronology for growth and geochemical tracers (Smithers and Woodroffe, 2001;Cole et al., 2000;Fleitmann et al., 2007).In Antongil Bay, the driest month on average in a given year is October (Kremen, 2003).Consequently B a/C a and G /B minima were assigned to October (Fig. 2), creating consistent age models for all cores.All years were then in terpolated linearly between the October anchor points using AnalySeries 2.0 (Paillard et al., 1996).This resulted in a time scale of monthly resolution with 1 2 equidistant data points.Annual anomalies were calculated by averaging all monthly values between January and December for any given year.This gave the best fit between cores, and allowed for direct comparison with modelled sediment yield data. Linear cor relations between proxy data were computed using the soft ware "R ".
The mean seasonal cycle of B a/C a, G /B , 51 8 Osw and 513C for each core were constructed by calculating the monthly mean over the 15yr record length.Comparing the mean seasonal cycle of G /B with the mean seasonal cycle of modelled river discharge for each of the three individual watersheds indicates that there is good agreement between the two data in terms of temporal alignment (Fig. 2).This gives us confidence in our age model for each of the coral core records.
Table 2. Correlation coefficients o f seasonal (upper) and annual average (lower) coral proxy data between the three coral sites M A SI, ANDRA and IFAHO, and LA-ICP-MS data for M A SI and MAS3.Significance levels are given in parentheses, and correlation coefficients, w hich are below the significance level of P < 0.05, are marked with a star (*).The annual average proxy values were calculated by averaging the months January to December for which the highest between core correlations were found. of the seasonal cycles and baseline averages of individual proxies.At the end of the results section, we present hydrological model data for river discharge and sediment yield to test how well corals directly record river runoff for the three rivers influencing our corals.

Seasonal and interannual variability o f Ba / Ca and G /B between cores associated with the Antainambalana watershed
For the common period of 1991 to 2005, interannual variations in B a /C a and G /B , for the cores M ASI and MAS3, were compared in order to test their reproducibil ity (Table 2).For B a/C a, we measured both cores with LA-ICP-MS (LA-) and replicated M ASI with solution ICP-MS on a monthly resolution.The mean B a /C a be tween cores and techniques differed, with MAS3 having the lowest baseline of 5.51 ± 0.71 pmol mol-1 .The mean B a /C a measured by LA-and solution ICP-MS for MASI were 6.05 ± 0.75 pmol mol-1 and 6.75 ± 0.78 pmol mol-1 , respectively, with standard deviations (la ) overlapping (Ta ble 3).The interannual variations between cores were highly reproducible and statistically significant (> 9 5 % ), with an nual mean M ASI and MAS3 LA-ICP-MS B a/C a sharing 78% (R = 0.88; P < 0.01) of the variance (Fig. 3).M ASI LA-and solution ICP-MS data shared 52 % (R = 0.72; P < 0 .0 1 ) of the variance.The interannual variations in G /B profiles for M ASI and MAS3 for the period 1991 to 2005 were also reproducible (Fig. 3; Table 2).Seasonally, the two cores shared 90% (R = 0.97; P < 0 .0 1 ) of the variance, and on annual means 45 % (R = 0.67; P < 0.01); all statistically significant at the 95% level.The mean G /B between M ASI (0.95 ± 0 .0 3 ) and MAS3 (0.92 ± 0 .0 3 ) differed, yet their standard deviations overlapped (Table 3).Similar to B a/C a, MAS3 showed the lowest baseline level in G /B .

Seawater analyses
We established the local salinity-518 O relationship from a set of water samples to subsequently reconstruct salinity from coral skeletal 51 8 Osw, as derived from coupled coral S r/C a and 5180 measurements (Ren et al., 2002).We obtained the seawater 5180 and salinity for seven water samples collected in October 2008, representing the dry season, and a further sample taken in March 2007, capturing the wet season signal.The seven samples were collected along an inshore-offshore gradient between the mouth of the river Antainambalana and the M ASI coral.The salinity of each water sample, taken in pre-cleaned 1 litre HDPE bottles, was measured with a hand held probe (Vernier) and a fraction was stored in air-tight 100 ml glass bottles.All water samples were subsequently poisoned with HgCU to prevent biological activity.Samples were then analysed for 5180 on a Thermo Finnigan Delta+ mass spectrometer equipped with a GASBENCH-II prepara tion device at the VU Amsterdam.A 0.5-1 ml water sample was injected through the septum cap of a 1 0 ml exetainer vial filled with a mixture of He and 0.2 % CO2 to equilibrate with the oxygen in the headspace CO2 for 24 h at 22 °C.Subse quently, the headspace mixture is transported by a He carrier flow (dehydrated using NAFION tubing) for analysis of CO2 in the mass spectrometer after gas separation in a GC col umn.Values are reported as 518 O vs. V-SMOW with a long term reproducibility better than 0.1 % o (1SD) for a routinely analysed lab water standard.

Results
The results section begins with a comparison of B a /C a and G /B variability between cores from a single site to assess proxy reproducibility.Next, we compare the variability of four proxies at three different locations to assess regional similarities.This section also includes a detailed description

Seasonal and interannual variability o f four river runoff proxies associated with three individual watersheds
Here, we compare G /B profiles and geochemical data for three coral cores associated with three individual watersheds across Antongil Bay.Geochemical data were measured on splits of the same powdered coral samples representing ap proximately 1 month of growth for individual cores.The three cores, M A SI, ANDRA and IFAHO, showed strong seasonal cycles in the monthly G /B time series for the 15 yr period studied, from January 1991 to December 2005 (Fig. 4).Similarly, strong seasonal cycles were observed in the monthly B a /C a time series of M ASI and IFAHO, yet the seasonal cycles of ANDRA were not as clearly defined (Fig. 4).We followed the method of Ren et al. (2002) to cal culate 51 8 Osw, which assumes that coral S r/C a is solely a function of SST and that coral S180 is a function of both SST and 51 8 Osw (Figs.A Í and A2).The 51 8 Osw and S13C seasonal cycles of M ASI were both strong and well de fined (Figs.A3 and A4).However, in ANDRA and IFAHO the 51 8 Osw and S13C seasonal cycles were less defined yet showed strong variability (Figs.A3 and A4).
For annual averages, luminescence (G /B ) showed the highest spatial correlations between the three cores/regions for the 15yr period out of all proxies studied (« = 15).AN DRA luminescence was statistically correlated with both M ASI (R = 0.69; P = 0.005) and IFAHO (R = 0.63; P = 0.013) (Table 2).The relationship between M ASI and IFAHO luminescence was just outside of the significance level (R = 0.50; P = 0.056).This relationship was signifi cant, however, when considering the seasonal data (Table 2).
Conversely, the strongest and only statistically signifi cant annual average B a/C a relationship was between cores M ASI and IFAHO (R = 0.66; P = 0.008) (Table 2).AN DRA B a /C a showed a weaker relationship with both MASI (R = 0.37; Z = 0.17) and IFAHO (R = 0.27; P = 0.32), which were not significant (Table 2).
The strongest relationship between average annual 51 8 Osw values was again between M ASI and IFAHO (R = 0.41; T3 = 0.13), yet not significant (Table 2).This relationship was significant, however, when considering the seasonal data (Table 2).The correlation coefficients of ANDRA 51 8 Osw with M ASI and IFAHO 51 8 Osw were lower and not signifi cant (Table 2).
The strongest relationship between average annual S13C values was between ANDRA and IFAHO (R = 0.32; P = 0.25), yet not significant (Table 2).The correlation coeffi cients of M ASI S13C with ANDRA and IFAHO S13C were both lower and also not significant (Table 2).

Luminescence (G /B ) seasonal cycles and baseline averages
Comparing the 15yr time series of all three G /B records revealed differences in baseline values and signal ampli tudes (Fig. 5a).M ASI G /B had the highest mean value and standard deviation (0.95 ± 0.034) compared to ANDRA (0.90± 0.029) and IFAHO (0.89 ± 0 .0 1 8 ) (Table 3).M ore over, ANDRA and IFAHO had similar mean values.How ever, the signal amplitude of ANDRA was far greater than IFAHO (Fig. 5a).During the dty season ANDRA G /B was below IFAHO G / B , yet during the wet season ANDRA G /B was higher.This is further highlighted by the mean seasonal cycles of G /B for the three cores (Fig. 6 ).Here, it becomes clear that the M ASI G /B signal was far higher than both the ANDRA and IFAHO G /B signals for the entire calendar year, while the ANDRA G /B signal overtakes the IFAHO G /B signal during the wet season.Standard deviations indi cate the variability of each site over the 15yr period (Fig. 6 ), with M ASI G /B overlapping ANDRA (not IFAHO) only in four out of twelve calendar months.Three of those four months were in the dty season.

B a /C a seasonal cycles and baseline averages
The most noticeable contrast between the G /B and B a /C a time series was in their respective baseline aver ages (Fig. 5b).Although the M ASI B a/C a mean value (6.75 ± 0.78 pmol mol-1 ) was highest of the three records, similar to the results of G /B , a significant difference be tween the mean B a/C a values of ANDRA and IFAHO was observed (Fig. 5b and Table 3).The mean IFAHO B a/C a value (5.42 ± 0.28 pmol mol-1 ) was significantly higher than ANDRA (3.80 ± 0 .5 8 pmol mol-1 ), despite the higher range and variability of ANDRA (Table 3).This contrast in results www.biogeosciences.net/9/3063/2012/Biogeosciences, 9, 3063-3081, 2012 is further emphasised when comparing the mean seasonal cy cles of the three B a /C a records (Fig. 6 ).Only in November and December was there an overlap in standard deviations between M ASI and IFAHO (Fig. 6 ).All ANDRA monthly Ba / Ca values were significantly below the Ba / Ca values of IFAHO.Further, a small decrease was observed in the B a/ Ca signal of ANDRA occurring between February and May (wet season).This was not the case for M ASI or IFAHO (Fig. 6 ).

á18Osw seasonal cycles and baseline averages
Measuring the salinity and S180 of seawater samples, we established a regional regression equation and applied it to transform coral derived S1 8 Osw values into reconstructed salinities (Fig. 7).The core with the most negative mean 51 8 Osw value (-0.50 ± 0.39 %o) was M ASI (Table 2).When converting the skeletal S1 8 Osw signal, the mean reconstructed salinity of M ASI equated to 25.54 ± 2.93 psu (Table 3).IFAHO had a mean S1 8 Osw value of -0.27 ± 0.20 %o, which equated to a mean reconstructed salinity of 27.26 ± 1.50 psu (Table 3).The most positive mean S1 8 Osw value was AN DRA (0.17±0.28% o), equating to the highest mean recon structed salinity of 30.55 ± 2.07 psu (Table 3).The standard deviations given for the mean S1 8 Osw seasonal cycles of M ASI and IFAHO overlap in every month with the exception of February (Fig. 6 ).Both showed a mean seasonal S1 8 Osw cycle decreasing during the wet season and increasing during the dty season (albeit IFAHO values start to increase earlier than M ASI) (Fig. 6 ).This is not the case for ANDRA, which showed a reverse mean seasonal cycle whereby S1 8 Osw val ues increased during the wet season and decreased during the dry season (Fig. 6 ).Moreover, during the months November to February there was no overlap of standard deviations be tween IFAHO and ANDRA as the 51 8 Osw values of IFAHO were significantly more negative than ANDRA (Fig. 6 ).

á13C seasonal cycles and baseline averages
Core ANDRA gave the most negative mean S13C signal, being -3.33± 0.65% o (Table 3).The core with the most positive mean S13C signal was IFAHO (-2.83 ± 0.45 %o), whereas M ASI measured -3.01±0.58%o (Table 3).All cores showed high variability over the 15yr period (Ta ble 3).Moreover, the monthly standard deviations given for  1990 1992 1994 1996 1998 2000 2002 2004 2006 2008  and Table 3).Further, the mean seasonal cycles were incon sistent between cores as both IFAFIO and M ASI showed a depleted S13C signal during the wet season and an enrich ment during the dty season (albeit IFAFIO again enriched earlier than M A SI), whereas, ANDRA showed a bimodal cycle (Fig. 6 ).

River discharge
The river with the highest modelled discharge was the An tainambalana, associated with coral M A SI, which had a mean discharge rate of 260 ± 92 m3 s_ 1 (Table 4).The rivers Ambanizana and Anaovandran, associated with corals AN DRA and IFAHO, respectively, had a statistically similar modelled discharge rate, however, of significantly lower magnitude to the river Antainambalana (Table 4).The mean discharge rate of the Anaovandran (8.2 ± 4.7 m3 s _1) was slightly higher than the Ambanizana (6 .8 ± 4.0 m3 s_1).This order of relative discharge rate between the three rivers was similar to the proxy results of G /B (Fig. 6 ).

Sediment runoff
The river with the highest modelled sediment runoff was the Antainambalana, associated with coral M A SI, which also recorded the highest B a/C a values (Fig. 6 b and Table 4).However, unlike the modelled discharge data, there was a sig nificant difference between the modelled sediment runoff of the two other rivers (   significantly higher than for the river Anaovandran, asso ciated with the coral IFAHO (Table 4).The relative order of sediment runoff rates between the rivers ANDRA and IFAHO is opposite to those of the Ba / Ca results, as ANDRA B a/C a values were lowest (Fig. 6 b).

Reproducibility o f Ba / Ca and G /B between cores (MASI and MAS3) from the same watershed
We demonstrate that two coral cores from the same water shed in Antongil Bay share a significant amount of interan nual variation in both B a /C a and G /B (Fig. 3).This indi cates that both of these terrestrial runoff proxies respond to a common environmental signal.These results are in agree ment with previous analyses, in which G /B signals were compared on longer time scales between the coral cores M ASI and MAS3 (Grove et al., 2010).Grove et al. (2010) used lOOyr of G /B data, which revealed an even higher cor relation than for the 15yr period considered here.A compar ison of Ba / Ca and G /B was only performed for core MAS 1, which showed a significant correlation for annual mean val ues (Grove et al., 2010).
Here, we replicated B a/C a for a second core (MAS3) and can confirm that interannual variations for this particular wa tershed are reproducible (Fig. 3).This agrees with similar studies from adjacent watersheds in the Great Barrier Reef (GBR), where B a/C a profiles showed significant correla tions between cores (Alibert et al., 2003).Slight offsets in mean B a /C a and G /B signals between M ASI and MAS3 are most likely related to different hydrodynamic regimes of flood plume currents between the reef slope (MASI) and reef flat (MAS3) sites along the Nosy Mangabe island fring ing reef.Similar small-scale differences in terrestrial runoff proxies have been observed for GBR catchments related to 'island w ake' effects (Jupiter et al., 2008;Lewis et al., 2011).
The differences in absolute values observed between LA-B a /C a and solution ICP-MS B a /C a in M ASI are likely related to time averaging in monthly milled samples com pared to sub-weekly LA-data (Fig. 3).However, the relative changes in B a/C a do share 52% of the variability between different techniques and the main sediment runoff spikes were reproducible, i.e. years 1991, 2000.

Spatial linkages between coral proxies o f terrestrial runoff across Antongil Bay
Proxy validation is a common problem in coral palaeoclimatology, as long-term in-situ data are rarely available (Jones et al., 2009).Lower resolution data are often applied as a substitute to calibrate proxies, i.e. satellite and model data (Corrège, 2006;Reynolds et al., 2002;Quartly et al., 2007).In this study, luminescence was the only proxy to show a considerable relationship between all three corals.For the majority of other proxies, vital effects and localised differ ences were likely more dominant than the regional climate signal expressed in the values.To best understand these dif ferences between river signals at individual corals and at their respective river mouths, we compared our proxy data with hydrological model data.Although not ideal, the model data give a good indication of both sediment yield and river dis charge for the three watersheds studied.A strong relationship was observed between annual av erage G /B of the three coral cores M A SI, ANDRA and IFAHO, suggesting that corals are recording a regional sig nal likely reflecting HA runoff.This argument is further strengthened when considering the modelled discharge data.Core M ASI and the river Antainambalana (associated with M ASI) showed highest G /B values and modelled discharge, respectively, compared to the other two cores/rivers.M ore over, corals ANDRA and IFAHO showed statistically sim i lar baseline averages in G /B , again replicated by the mod elled discharge data.As HA runoff is linked to river dis charge (Lough, 2011a;Grove et al., 2010), this likely ex plains the patterns observed in our corals.The only contrast between the two datasets was that ANDRA G /B peaked above IFAHO G /B during the wet season.Discharge data suggest otherwise, whereby the mean, range, maximum and minimum river discharge of Ambanizana, associated with coral ANDRA, are all less than the Anaovandran, associated with the coral IFAHO.At this stage, we cannot exclude the possibility that the recorded discharge signal at coral AN DRA may also be influenced by adjacent rivers.Also, in con trast to the large watershed of the Antainambalana, the rivers influencing ANDRA and IFAHO have much smaller water sheds, which increases uncertainties in modelled discharge due to the relatively large model grid size (50 km).
W hen considering both the 51 8 Osw baseline averages and mean seasonal cycles, it becomes clear that strong hy drographic differences exist between the three coral sites.Coral M ASI recorded the freshest waters, comparable to the modelled discharge rates for the Antainambalana.However, the reconstructed salinity signal at ANDRA indicated most saline conditions as well as a slightly reversed mean sea sonal water cycle (Fig. 6 c).As both the river Ambanizana and Anaovandran, associated with ANDRA and IFAHO re spectively, have similar watershed sizes and modelled dis charge rates, the 51 8 Osw baseline averages and seasonal cy cles were expected to be similar.Such inconsistencies are likely a result of a difference between the distances from the river mouth to the coral.
Coral ANDRA is located 7 km from the Ambanizana river mouth, compared to IFAHO, which lies 4.5 km from the Anaovandran.The river signal at ANDRA is therefore likely to have been diluted by seawater via conservative mixing more than the signal at IFAHO, giving it a higher recorded salinity.Alternatively, the higher salinity signal at ANDRA may be linked to currents.As coral ANDRA is located fur ther from the river mouth than IFAHO, it is increasingly likely that currents may channel the freshwater signal, asso ciated with the river Ambanizana, away from the coral head.Nevertheless, according to proxy data, coral ANDRA re ceives no freshwater signal during the warm/wet season, and furthermore, has a slight increase in salinity.This increase might be related to the S1 8 0 hydrological balance (evapora tion) of the water body influencing ANDRA at this time.
Comparing the B a/C a signals between cores and with modelled runoff data augments the argument that coral AN DRA is not influenced by runoff from the river Ambanizana.Coral M ASI showed the highest recorded levels of B a/C a compared to the other two corals, which is in agreement with the modelled data, as the river Antainambalana had the high est modelled sediment runoff.As the B a /C a annual averages of M ASI and IFAHO are statistically correlated, they seem to be recording a regional sediment runoff signal (Sinclair and McCulloch, 2004;Alibert et al., 2003;McCulloch et al., 2003;Fleitmann et al., 2007).Coral IFAHO showed a sig nificantly lower mean B a/C a signal.This signifies that the source input at the river Anaovandran is considerably lower than the river Antainambalana, which is again in agreement with modelled sediment runoff data.
Model data for the river Ambanizana, associated with coral ANDRA, indicates that sediment runoff is far higher than that of the river Anaovandran, associated with IFAHO.ANDRA resides a further 2.5 km from its river source than IFAHO, therefore any river signal from the Ambanizana would be diluted/mixed by seawater for an extra 2.5 km.However, given the significantly higher modelled sediment runoff for the Ambanizana, it is expected that ANDRA B a /C a would still be higher or similar to IFAHO.This is not the case, as the B a/C a signal in ANDRA is significantly less than IFAHO.Given that the annual average B a /C a results of ANDRA do not correlate with either M ASI or IFAHO, it is therefore likely that (1) skeletal B a/C a variability is linked to a different source other than sediment runoff, (2 ) skele tal B a /C a variability is derived from the largest watershed and its Ba / Ca signature is well mixed and diminished before arriving at the coral site, and/or (3) low sediment runoff is associated with the river Ambanizana watershed, contrary to what modelled data suggest.
Although ANDRA and IFAHO catchments have similar sizes and draw from the same mountains covered in dense rainforest, differences in hinterland topography and vegeta tion cover prevail.Such differences should also be consid ered when explaining the patterns observed between Ba and HA.The ANDRA river watershed follows a valley with a steep slope covered with rainforest (Figs. 1 and 3 in Kremen et al., 1999).The river associated with IFAHO drains from a shallower slope and eventually runs through a plain with less vegetation cover (Windley et al., 1994 and references therein;Figs. 1 and 3 in Kremen et al., 1999).This shallow slope allowed for easy human settlement and subsequent de forestation (Figs. 1 and 3 in Kremen et al., 1999;Green and Sussman, 1990).The difference in slope influences the flow www.biogeosciences.net/9/3063/2012/Biogeosciences, 9, 3063-3081, 2012 speed and thus the amount and type of dissolved and par ticulate transports (Milliman and Syvtski, 1992;Larsen and Webb, 2009).The continuous vegetation cover at ANDRA binds and protects sediment from erosion but allows leaching of HA, while the plain at IFAHO is much less densely veg etated, most likely facilitating sediment erosion (Douglas, 1967).HA input from the river source associated with AN DRA would therefore be higher than that of IFAHO, as more dissolved HA are leached from a watershed with a denser vegetation cover, thus flushing more dissolved HA down the steep gradients into the marine system.As the vegetation re tains sediments, this might also partly explain why Ba levels at ANDRA are low.However, modelled sediment yield sug gest otherwise, which indeed takes factors such as elevation, slope, soils, etc. into account.
Previous studies have also found no relationship between B a/C a and river discharge across a water quality gradient (Jupiter et al., 2008;Prouty et al., 2010;Lewis et al., 2011).Again, in contrast to the large watershed of the Antainambal ana, the rivers influencing ANDRA and IFAHO have much smaller watersheds, which increases uncertainties in mod elled sediment yield due to the large model grid size.Fur thermore, the model data indicate the sediment yield at the river mouth whereas the corals record the ambient B a/C a signals (not sediment) at the reef site along a water quality gradient (Lewis et al., 2011).
The low B a /C a values and high reconstructed salinities of ANDRA provide evidence that the river Ambanizana is only marginally influencing the coral site.Yet, the high G /B values suggest otherwise, indicating that a runoff signal does exist during the wet season.Indeed, the strongest annual av erage G /B relationship observed between cores was between coral M ASI and ANDRA.Therefore, it is more likely that the HA signal reaching coral ANDRA is a mixture of signals originating from the river Antainambalana, associated with M A SI, and the river Ambanizana, associated with ANDRA.This is plausible given the clockwise direction of the cur rents within the bay (Fig. 1), the high concentrations of HA associated with each watershed and the conservative mixing nature of HA (Bowers and Brett, 2008).Unlike HA, barium behaves non-conservatively in estuaries, as it is influenced by processes such as phytoplankton cycling (Sinclair, 2005;Hanor and Chan, 1977;Coffey et al., 1997).Therefore, the riverine Ba signal associated with M ASI may well have di minished by the time it reached coral ANDRA.Further, the salinity signal (51 8 Osw) from the Antainambalana may have been lost due to mixing of different water masses (Fig. 7).As HA are conservative and are only associated with terrestrial inputs, they are duly transported to and recorded by coral ANDRA.
The skeletal S13C signals of all three cores showed sim i lar baseline averages and high standard deviations, making it difficult to statistically differentiate between signals.Coral S13C can be affected by a number of vital effects includ ing kinetic effects, light, pH variation at sites of calcifica tion, skeletal architecture and the influence of metabolic CO2 (McConnaughey, 2003;Felis et al., 2003;Rollion-Bard et al., 2003), which have likely influenced the S13C variability within and between cores.Nevertheless, the mean seasonal cycles of M ASI and IFAHO showed a depletion during the wet season.This probably reflects either a depleted S13C sig nal of DIC associated with the river plume (Moyer, 2008;von Fischer and Tieszen, 1995;Swart et al., 1996;Marin-Spiotta et al., 2008;Moyer and Grottoli, 2011), or a decrease in pho tosynthesis reducing the depletion of 12C in the carbon pool (Grottoli, 2002;Grottoli and Wellington, 1999;Weil et al., 1981;Swart et al., 1996;Reynaud-Vaganay et al., 2001;Rey naud et al., 2002).Both would yield an inverse relationship with increasing runoff.As the freshest waters are associated with coral M A SI, this might explain the more depleted S13C values compared to IFAHO.
The S13C signal of coral ANDRA was more depleted than both M ASI and IFAHO.This is surprising as it was not in fluenced by discharge directly, and therefore values were ex pected to be more enriched (Moyer, 2008;von Fischer and Tieszen, 1995;Swart et al., 1996;Marin-Spiotta et al., 2008;Moyer and Grottoli, 2011).Further, the mean seasonal cy cle of ANDRA S13C was bimodal, as both S13C enrichment occurred during the peak runoff season as well as the dty season.As ANDRA is not (or only marginally) influenced by runoff, there are likely many other factors contributing to skeletal S13C variability and the baseline average, including ambient seawater productivity.Interestingly, the mean sea sonal cycle of B a/C a in ANDRA also showed a decrease in March when S13C enriched slightly (Fig. 6 ).This may be linked to phytoplankton uptake of Ba during peak runoff, when nutrients are plentiful, causing enrichment of S13C due to the preferential uptake of 12C by phytoplankton (Sinclair, 2005 and references therein;Stecher and Kogut, 1999).An increase in primary production causes Ba to be scavenged from the water column due to the active cycling of algal blooms (Sinclair, 2005;Stecher and Kogut, 1999).As the nutrient levels are reduced, decaying algae will increase Ba concentrations within the water column by recycling.Cor relating the annual average B a/C a values with the annual average S13C values indeed gives an indication that this is the case.Although not statistically significant, ANDRA gave the highest correlation (R = 0.46; P = 0.086), compared to M ASI (R = 0.029) and IFAHO (R = 0.13).
In this study we demonstrate that B a/C a and G /B sig nals from the same watershed are reproducible.However, strong localised signals are observed between cores associ ated with different watersheds due to the large distances sep arating the corals.G /B was the only proxy which showed a regional similarity across the bay, although the relationship between M ASI and IFAHO was just outside the 5 % signif icance level.A strong relationship was observed in B a/C a for M ASI and IFAHO, yet not with ANDRA.This is likely related to ANDRA residing a further 2.5 km from its as sociated river mouth compared to IFAHO and topographic and vegetation cover differences between watersheds, despite having a similar sizes.No significant regional signal was observed for S1 8 Osw and S1 3 C.The S13C signal was likely overwhelmed by coral vital effects and in-situ productivity, whereas the S1 8 Osw signals are likely inconsistent due to a combination of vital effects and site-specific environmental differences in the hydrological balance.
Discrepancies between the proxy results and modelled data indicate that corals are not ideally suited for directly comparing river systems.Absolute proxy values give an in dication of the ambient concentration surrounding the coral at the time of precipitation, yet not the source input.Prouty et al. (2010) compare B a /C a baselines for Hawaiian corals to other published B a /C a records from the GBR (McCulloch et al., 2003) and Kenya (Fleitmann et al., 2007), relating differ ences to river input.This is problematic since the Ba / Ca sig nal at each site reflects both the distance from the source and the source input.Moreover, mixing gradients (distance), cur rents, proxy behaviour and vital effects can all influence the precipitated skeletal signal significantly from the moment it leaves the river source.In a recent study, Tewis et al. (2011) provided evidence of the impact small-scale hydrodynamic differences have on skeletal B a /C a across the complex hy drography of reefs in the GBR from multiple cores across a water quality gradient.However, by combining runoff prox ies and comparing baseline averages and mean seasonal cy cles, aided by hydrological model data, a good overview of the runoff products influencing the coral sites was achieved in our study.Moreover, all proxies provide information on the runoff dynamics of the bay system, which will assist both terrestrial and marine management programmes in M adagas car.
an ideal proxy for paleoclimate reconstructions and spatial comparisons of corals (not rivers).Yet, the skeletal S1 8 Osw signal showed little consensus between cores, which was likely due to vital effects, local differences in the hydrologi cal balance and their vicinity to the river mouths.Other physiochemical parameters such as currents also need to be con sidered when interpreting proxy results.Even though a coral may reside closer to one river than another, currents can de termine which river signal the coral receives.Nevertheless, comparing proxy baseline averages and mean seasonal cy cles provides a good overview of the runoff dynamics over a bay system.

Conclusions
Models provide an estimate of discharge and sediment runoff at the mouth of a river, whereas corals record a signal po tentially several kilometres away.Subsequently, the riverine signal can be modified by the time it reaches the coral.The distance from the river source is a primary reason why com paring absolute runoff proxy values to differentiate between watersheds is unreliable.The further a coral resides from the river mouth, the more the proxy signal is mixed/diluted by seawater.Depending on the behaviour of the proxy, the dis tance from the river can also have dramatic effects on the recorded signal at the coral.Proxies such as B a /C a and S13C are non-conservative mixers, and therefore altered by recy cling processes such as phytoplankton uptake.Nevertheless, we find B a /C a to be reproducible within the same watershed and also across distant watersheds given there is a close prox imity of the coral to a river mouth and/or associated with a large river.Humic acids and the S180 signal mix conserva tively in the water column.Indeed, there was a good regional relationship observed between coral G /B signals, making it thus altering the amount and characteristics Published by Copernicus Publications on behalf o f the European Geosciences Union.C. A. Grove et al.: Spatial linkages between coral proxies o f terrestrial runoff

Fig. 1 .
Fig. 1.Map o f Antongil Bay in NE M adagascar showing the three coral sites (stars) and closest rivers with respective water sheds (grey shaded areas).Rivers are marked by the numbers 1 (M ASI, Antainambalana River), 2 (ANDRA, Ambanizana River) and 3 (IFAHO, Anaovandran River).The bay circulation is marked by an arrow indicating a clockwise direction.The location o f towns (black circles) and national park boundaries (Makira-Masoala Pro tected Landscape: think grey lines) are marked on the map accord ingly-

Fig. 2 .
Fig. 2. Average seasonal cycle o f SST (a) (ERSST; Smith et al.. 2008) for Antongil Bay (52° E. 16° S) calculated by averaging ev ery month over a 15 yr period (1991-2005).Similarly, over the same 15yr period, the average seasonal G /B (solid line) and modelled river discharge data (dashed line) are calculated for the corals (b) M A SI, (c) ANDRA and (d) IFAE10 and their closest respective rivers (b) the Antainambalana River, (c) the Ambanizana River and (d) the Anaovandran River.

Fig. 3 .
Fig. 3.The three monthly B a /C a (left) profiles and two G /B (right) profiles between January 1991 and December 2005, signifying proxy reproducibility in both seasonality and interannual variability.The three B a /C a profiles (a and b) include LA-M A SI (solid).LA-MAS3 (dotted) and solution ICP-MS M A SI (dashed).The two G /B profiles (c and d) include M A SI (solid) and MAS3 (dotted).Absolute values are shown for B a /C a and G /B in the upper plots (a and c. respectively), and annual anomalies for both B a /C a and G /B are shown in the lower plots (b and d. respectively).

Fig. 4 .
Fig. 4. M onthly G /B (a.b and c) and B a /C a (d.e and f) time-series over the 15 yr period between 1991 and 2005 for the three corals M ASI (a and d).ANDRA (b and e) and IFAHO (c and f).The annual average (dashed line) G /B (a.b and c) and B a /C a (d.e and f) values were calculated by averaging the months January to December for each o f the 15 yr.

Fig. 5 .
Fig. 5. Cross-core comparison o f monthly G /B (a) and B a/ Ca (b) absolute values for the complete 15 yr time-series o f the three cores M A SI (solid), ANDRA (dotted) and IFAHO (dashed).
mean seasonal cycles indicate that individual S13C baseline values were statistically similar over the 15 yr (Fig.6

Fig. 6 .
Fig. 6.M onthly averaged G /B (a).B a /C a (b).á 1 8 Osw (c) and <î13C (d) for M A SI (solid circles).ANDRA (open circles) and IFAHO (triangles), indicating the average seasonal cycles.Monthly averages were calculated for the 15 yr period spanning January 1991 till December 2005.The standard deviations for individual months are given as error bars.The á 1 8 Osw and á 13C is given relative to V-PDB.

Fig. 7 .
Fig. 7. Calibration of the <518 0 of water samples with salinity m ea surements used to reconstruct salinity from coral <5180 Sw-All data points are given as averages o f two samples taken during the dry season (solid circles) and wet season (open circle).The standard deviation is given as error bars.

Fig
Fig. AÍ.The S r/ Ca time-series o f coral cores M A SI (a), ANDRA (b) and IFAHO (c) for the common period 1991-2005.

Table 1 .
Coral cores with GPS co-ordinates, growth rates, total core length, distance and name to the closest river source.