Evaluating the effect of data-richness and model complexity in the prediction of coastal sediment loading in Solomon Islands

This study applied two catchment sediment export models, using input sources of varying data-richness, to explore seven sediment export scenarios. Catchment sediment export estimations and the frequency of total suspended solids (TSS) exceedance were compared using in situ stream observations and water quality guideline thresholds. Simulation results from the catchment models were used as inputs to three coastal sedimentation models. Outputs from the sedimentation models were evaluated using coastal sedimentation measurements under both data-poor and data-rich catchment sediment export scenarios.

The study site is the approximately 75 km² Jejevo catchment (−8.1° S, 159.12° E) located on the windward side of Isabel Island in Isabel Province, Solomon Islands (figure 1). Solomon Islands is a country formed by over 900 islands, where six of the largest islands form the majority of the approximately 650 000 population [43]. Annual rainfall total averages measured in coastal areas of Solomon Islands have mostly been in the range between 3 000 mm and 5 000 mm. Although no long-term rainfall stations exist at higher elevations relative to the coast, it is thought that local topographies may increase the average annual rainfall total to the order of 9 000 mm in some places [44]. While nearby catchments have been disturbed by logging operations, the study area remains a predominantly undisturbed from human interference (a small gardening settlement exists near the mouth of the river) with prevailing steep tropical rainforest (where erosion processes are not well-known). A rare high-quality dataset was collected in the study area, providing a unique opportunity to investigate the value of increased investment in data-richness over a range of model complexities. Hourly in situ stream monitoring and flood event sampling (rising stage samplers and ISCO automated samplers) collected over the duration of 2013, and coastal sedimentation observations over this same period were used from the downstream monitoring site (due to continuous data availability and quality considerations). There were also hourly recorded local weather (rainfall, wind, relative humidity, air temperature, solar radiation, and barometric pressure) observations (Nuha Camp Weather Station) available during this time (figure S1 (available online at stacks.iop.org/ERL/15/124044/mmedia)).

Zoom In Zoom Out Reset image size

Event sampling and analysis were combined with in situ turbidity from data loggers (YSI 6820 V2) deployed in the Jejevo stream to determine the approximate relationship with the TSS measured in the corresponding samples (figure S2). The best fit was found at approximately one NTU for every 2 mg l⁻¹ of suspended sediment in the stream (R² of 0.77 with events measured up to 227 NTU and 420 mg l⁻¹ using 40 samples; figure S3 and S4).

TSS frequency of exceedance was calculated for the observations and higher complexity catchment model development scenarios, drawing comparison to turbidity guidelines for thresholds in drinking and environmental water quality [45, 46]. Further evaluation was also able to be undertaken in the receiving waters by using a combination of settling tube deployments and sediment core dating. A sediment core taken from the nearshore site was found to show a sedimentation rate of 1.07 cm yr⁻¹ over the previous 40 year period [47]. This is equivalent to a sedimentation rate of 7.77 mg cm⁻² d⁻¹ adjacent to the Jejevo river mouth (assuming a sediment density of 2 650 kg m⁻³).

Two commonly used catchment models for sediment export estimations were selected to represent the lower and upper ranges of model complexity. The Integrated Valuation of Ecosystem Services Sediment Delivery Ratio Model (InVEST SDR version 3.2) [48] was selected as a low complexity model. InVEST is often applied where local data availability has been considered poor or non-existent [49, 50]. This modelling approach uses a digital elevation model (DEM) to estimate the yield of sediment from the revised universal soil loss equation with a ratio determining the proportion of soil loss exported from a respective stream [48]. Beyond the 30 m resolution DEM provided by the shuttle radar topography mission (SRTM) [51], spatial data is required to delineate the catchment (figure S5) and represent the distribution of soil types (from surveying [52]) and land use (derived from satellite analysis of tree coverage [53] with a 10% tree canopy cover threshold for deforestation [54] determining 99.98% of the catchment to be covered by rainforest). Furthermore, two rainfall erosivity (MJ mm ha⁻¹ hr⁻¹) equations using rainfall in mm were compared as a current constraint is that most studies in soil erosion have been done in agricultural landscapes of the United States of America [55, 56]. The Roose equation

$$\begin{equation*}R = MEAN\,ANNUAL\,RAINFALL \times MULTIPLIER\end{equation*}$$

is recommended for InVEST with a multiplier of 0.2–0.6 expected for mountainous tropical areas near the ocean [57]. As with a previous study in the region [49], 0.5 was selected as the multiplier as a compromise for most cases. The second equation compared in this study is the Bols equation [58]

$$\begin{align*} & R =\nonumber\\& \small \frac{{2.5\! \!\!\times\!\!\! {{\left(\! {MEAN\ MONTHLY\ RAINFALL} \!\right)\!}^2}}}{{100 \!\!\!\times\!\!\! \left(\! {0.073 \!\!\times\!\! \left(\! {MEAN\ MONTHLY\ RAINFALL} \!\right) \!\!+\!\! 0.073} \!\right)}}\end{align*}$$

which has also been used in previous work [59, 60].

The Soil and Water Assessment Tool (SWAT) [61] was selected as a similar spatially distributed catchment model that has been widely applied [13, 61, 62], and with a sufficiently higher complexity for comparison in this study. The data inputs for the different model development scenarios are detailed in table 1. Scenario 7, representing the most data-rich model, is the only gauged catchment scenario and involved calibration to in situ data and event samples. Under this condition, calibration to streamflow (Redback current meter, INW pressure transducer, Sontek Argonaut SL1500, Sontek FlowTracker, and Sontek RiverSurveyor) was first performed, followed by calibration to TSS using the best fit observed TSS to turbidity ratio on the logged turbidity measurements (table S1). Calibration was carried out using sequential uncertainty fitting [63]. A normalised root mean squared logarithmic error (NRMSLE) was calculated for the sediment export estimated by each scenario relative to the logarithmic median observed in the catchment,

$$\begin{align*}NRMSLE = \frac{{\sqrt {{{\left( {{\text{log}}\left( {SCENARIO + 1} \right) - {\text{log}}\left( {MEDIAN\ OBSERVED + 1} \right)} \right)}^2}} }}{{{\text{log}}\left( {MEDIAN\ OBSERVED + 1} \right)}}\end{align*}$$

Three modelling approaches of varying complexity were used to quantify coastal receiving water sedimentation estimations. The sediment export values from scenario 3 (data-poor) and scenario 7 (data-rich) were both used as inputs for each sedimentation model (table 2). The simplest of the approaches considered was a one-dimensional basin sedimentation model (BSM) [64] that was derived as the centerline of the plume exiting the river mouth at a specified width, depth, velocity, and concentration. Beyond these inputs to this model, only one parameter is required to represent the expected 'removal rate' from the plume as it settles out of suspension selected at 0.0001 s⁻¹. Sensitivity analysis showed no significant change in the overall magnitude of the results with changes in removal rate. The BSM estimations for the sedimentation along the plume centerline were calculated on a 50 m grid using the distance from the mouth of the river.

Another approach applied in this study is sediment extent modelling (SEM) [36]. The SEM approach uses globally available surface current modelling [33] and a particle settling rate which is modified to be reduced in shallower depths than 10 m

$$\begin{equation*}REVISED\ SETTLING\ RATE = \frac{{SETTLING\ RATE}}{{{{10}^{{{0.5}^{0.1 \times DEPTH - 1}}}}}}\end{equation*}$$

The settling rate is a key parameter representative of the particle class size of sediment to be modelled, and was selected at 0.00005 m s⁻¹ based on coastal environment fine sediments in the concentration range observed [65]. This revised settling rate compensates for near-shore processes such as wave turbulence which may hinder or reduce the effective settling rate of a particle in suspension. Thereafter, a path distance algorithm is applied using a horizontal relative moving angle and the pre-defined 'forward' horizontal factors using a cost dataset comprised of the calculated seconds needed to traverse a single element of the grid using the surface velocities [36]. Sediment extent areas were calculated for each month's mean surface currents [33] using a surveyed bathymetry dataset (figure S6) and uniformly loaded from the source catchment, with aggregation of these monthly plume areas over the year. The bathymetry survey dataset applied to the SEM in this study was of higher resolution (sampled to 50 m from a higher resolution survey) to those applied in previous studies using global-scale coarse resolution sources on the order of 1 km with simulations carried out at 90 m resolution [36].

A Delft3D model [66] with a 50 m horizontal grid resolution of the receiving waters was developed as the highest model complexity scenario considered in this study. This scenario represents the development of a resource-intensive three-dimensional hydrodynamic, cohesive sediment transport [67], and ocean heat flux model [68] typical of the highest degree of complexity employed to model a coastal receiving water environment [69]. Eleven sigma-coordinate vertical layers were used in the model with higher resolution at the top of the water column reaching 0.17% of the total depth and down to one-third of the total depth at the bottom of the water column. Sedimentation was simulated over the full 2013 study period using the catchment sediment export loading from scenario 3 and 7. The tidal forcing signal was developed using astronomical constituents [70], while Mannings bed roughness was specified at 0.025 in the receiving waters, enhanced to 0.1 in reef regions [71], and 0.05 used in the lower Jejevo stream to the river mouth [72]. The bathymetry of the model was smoothed towards the open ocean boundary for numerical stability in the regions where steep bed slopes were present.

An averaged NRMSLE was calculated between the measured sedimentation rates during the 2013 study period, and the simulated sedimentation rates estimated by each of the coastal models. This performance measure was selected to fairly represent the relative error in the model domain over the larger scales of variation across the study domain and simulation results.

In all metrics (R², NSE, and PBIAS) the SWAT model calibration in scenario 7 demonstrated a 'very good' performance for streamflow and 'not satisfactory' performance for TSS [73], highlighting a much larger challenge in calibrating sediment export processes at a daily timescale (table S2). The observed mean TSS for the 2013 study period was 110 mg l⁻¹ (20 mg l⁻¹ geometric mean) with a standard deviation of 370 mg l⁻¹, whilst the calibrated SWAT model in scenario 7 achieved a mean TSS of 98 mg l⁻¹ (0.5 mg l⁻¹ geometric mean) with a standard deviation of 312 mg l⁻¹ over the same period. Minimums between the observed and simulated datasets were more similar with 0 mg l⁻¹ for scenario 7 and 2 mg l⁻¹ for the field measurements. In the maximum observations recorded during the 2013 study period, the peak range of the turbidity logger was reached for between 1 day and 1 week. The corresponding potential magnitude of sediment loading to events larger than the measurement range of the turbidity sensor were subsequently unavailable, and hence the maximum observed TSS was 2593 mg l⁻¹, whilst scenario 7 produced an estimated unrestricted maximum of 3305 mg l⁻¹ based on the available measurements.

The total estimated sediment export for the 2013 study period out of the Jejevo estuary was summed for each catchment model scenario and compared with the sediment export distribution observed from field monitoring in the lower Jejevo river (figure 2(A)). Between the two simpler InVEST modelling scenarios, the Roose rainfall erosivity equation was found to be an order of magnitude closer to the observed range compared to the Bols rainfall erosivity variant. In contrast, the SWAT model scenario using global data (scenario 3) was on the same order of magnitude as the Roose rainfall erosivity approach (scenario 1) with an error between the two InVEST scenarios. This result suggests that increasing model complexity without any investment in increased data-richness was of no increased value for predicting sediment export from this catchment in the study period.

Zoom In Zoom Out Reset image size

Increasing the rainfall data-richness from a global coarse resolution monthly source to global coarse resolution daily sources both improved the estimation of total sediment export to within one order of magnitude at 87.2% NRMSLE (scenario 4) and slightly worsened the estimation of total sediment export at 99.7% NRMSLE (scenario 5). Noting that both the CFSv2 global forecast model rainfall predictions and TRMM satellite remotely sensed rainfall estimations resulted in under predictions. Figure 2(B) presents the cumulative rainfall for each of the rainfall datasets used across the seven catchment scenarios. Whilst the rainfall total for the global monthly WorldClimv2 is higher than those for both the CFSv2 forecast model and remote TRMM satellite-based rainfall estimations, the sediment export predictions from the catchment under the CFSv2 rainfall dataset is increased and approaches the observed range. Results highlight the significance of capturing local rainfall events on a daily scale compared to monthly timescales when estimating sediment export from this catchment. Scenario 6 using local rainfall measurements with a much higher cumulative rainfall total, was the only scenario to overpredict the observed range with a 72.1% NRMSLE, however it was very close to the observed range. Again, this reinforces the importance of capturing the local rainfall events. As expected, the results of scenario 7 provided the only estimate in this study that matched the observed range, however the calibration of the more complex model comes with the introduction of parameter identifiability challenges [74]. This is especially concerning where differences in rainfall data have been found to have a significant impact on the resulting sediment export. In an environment where local measurement location bias can be very large within the spatial heterogeneity characteristic of local rainfall patterns at the catchment scale, significant recalibration effort is required [9].

The TSS exceedance time-frequency distribution revealed strong differences in the data-richness dimension across the SWAT model development scenarios 3–7 (figure 3). The ungauged scenarios (3–6) performed particularly poorly in the estimation of TSS exceedance for both drinking [45] and tropical estuarine environmental turbidity guidelines [46]. Whilst the data-poor scenario (scenario 3) predicted guidelines would not be exceeded at all in the study period, investing further in data-richness was able to predict that the guidelines would be exceeded. However, this revealed that exceedances were not predicted to occur with the same frequency as was observed in the catchment. Specifically, observed exceedance of environmental and drinking water guidelines occurring 23% and 60% of the time, respectively. Scenario 4 revealed a 7% frequency of drinking water guideline exceedance, whilst the environmental guideline was rarely exceeded at <1% frequency. The environmental guideline was not exceeded at all in scenario 5, with the drinking water guideline exceeded at a frequency of just 2%. The addition of local rainfall data in scenario 6 improved this estimate to 20% frequency for exceeding the drinking water guideline (40% underestimated), and 9% for the environmental guideline (14% underestimated). Scenario 7 provided a much-improved estimate of the TSS exceedance frequency for the environmental guideline threshold virtually matching the observed frequency of exceedance. However, there still remained more than a 20% underestimation in frequency of exceeding the drinking water guideline. In particular, the results from scenario 7 suggest there are still substantial challenges in simulating low-flow, low-turbidity patterns.

Zoom In Zoom Out Reset image size

Simulated estimates of coastal sedimentation rates were highly variable depending on data-richness and model complexity. Plume centerline sedimentation rate estimations from the BSM were plotted over the distance from the mouth of the Jejevo river (figures 4(E) and (F)). The nature of the method predicts a radial pattern of potential sedimentation impact, underpredicting the observed rates by a mean NRMSLE of 88.8% in the data-poor catchment export (scenario 3) and overpredicting with a mean NRMSLE of 421.9% in the data-rich catchment export (scenario 7). It is important to note that whilst the data-poor scenario has resulted in a lower error, correct interpretation of the results of the BSM would likely find the data-rich scenario to be the preferable result representing the maximum expected sedimentation from the plume centerline. Although when compared to the highest investment model development scenario (figure 4(B)), the sedimentation patterns and radial rates of the data-rich BSM depict a very different representation of sedimentation in the receiving waters.

Zoom In Zoom Out Reset image size

Increasing model complexity to the SEM saw an increase in error compared to the BSM whilst using the scenario 3 catchment sediment export (figure 4(C)). However, when the catchment sediment export from scenario 7 was used the mean NRMSLE was reduced by 46.2% (figure 4(D)). Sedimentation patterns from the SEM over the study domain allow the influence of the local geometry of the receiving waters environment to be examined, however results show no resemblance to the patterns predicted by the more detailed scenario (figure 4(B)).

Increasing the model complexity further (D3D model) without any added investment in data-richness (figure 4(A)) was found to further increase the NRMSLE of the resulting sedimentation rate predictions. This result indicated a negative return on investment of model complexity resourcing without any investment to improve data-richness. The D3D scenario representing the most investment in model complexity and data-richness in this study (figure 4(B)) has the lowest error of all the scenarios considered. Estimation of the potential impact from sediment loading into the nearshore was therefore found to highly benefit from the collection of in situ data.