Presentation Information
[AOS27-P05]Validation of an SPSS Estimation Model for Assessing Red-Soil Runoff Impacts in Coastal Okinawa: Generalization Performance and Parameter Site-Specificity
*Nakano Yuta1, Kono Kai1, Tokunaga Daisuke1, Kawata Hiroaki1 (1.NTT, Inc.)
Okinawa Prefecture possesses rich coastal waters with extensive coral reefs, serving as vital assets for residents' livelihoods including fisheries and tourism. However, massive red soil runoff from terrestrial areas has seriously impacted the coastal environment. Although runoff from development projects has declined since the Red Soil Runoff Prevention Ordinance of 1994, agricultural land still accounts for approximately 80% of total runoff [4].
Suspended Particle content in bottom Sediment per unit volume (SPSS) is a representative indicator for assessing red soil impacts, defined as the mass of soil particles smaller than silt per unit volume of bottom sediment (kg/m³) [1]. Administrative monitoring evaluates sedimentation on a nine-level ranking scale based on SPSS to set priority measures and reduction targets [5]. However, SPSS measurement requires sediment sampling and laboratory analysis at high cost; observations by Okinawa Prefecture are limited to approximately twice per year [1].
The red soil runoff process progresses from rainfall onset through deposition, dispersion, and recovery over roughly one day to one week. As Ohmija [1] demonstrated, SPSS generally peaks after the rainy season and decreases where strong typhoon- or seasonal-wind-driven waves resuspend deposited sediment, exhibiting distinct seasonal variations. Biannual observations cannot resolve such dynamics, necessitating higher temporal resolution estimation.
One approach is physical model-based prediction. Sakai and Nakandakari [2] proposed a concise sediment budget model: terrestrial red soil inflow as a positive factor, wave resuspension and tidal flushing as negative factors, sequentially updated by rainfall, wave, and tidal forcing. Physical models decompose individual processes in an interpretable manner applicable to countermeasure evaluation. However, under sparse observations, unrepresented processes, data constraints, and inter-site environmental differences manifest as estimation errors.
In this study, the model of [2] was applied to SPSS data from observation sites in the FY2022 Red Soil Runoff Prevention Verification Project [3], combined with Japan Meteorological Agency data. Each site was paired with its nearest meteorological station for forcing inputs.
Comparing baseline models (BL1: Persistent; BL2: Linear regression) with the proposed model (A1), BL2 achieved the best simple RMSE, but A1 performed best under Log-RMSE. Given that SPSS spans a wide range following a log-normal distribution, A1's Log-RMSE superiority suggests relatively uniform accuracy across both low and high value ranges.
We then examined performance when parameters set uniformly across all sites in [2] were optimized per site. For the river sediment inflow coefficient A, site-specific optimization improved accuracy, whereas for the wave resuspension coefficient M, accuracy remained virtually unchanged with a uniform value. This reflects that red soil inflow depends strongly on geographical factors (catchment area, land use, river morphology) with large inter-site variability, while wave resuspension is governed by relatively homogeneous mechanics. Accurate characterization of site-specific river inflow properties is thus key to SPSS estimation.
These results demonstrate that site-specific river inflow parameter calibration directly improves physical model-based SPSS estimation, and that broad-area application is feasible with minimal parameter adjustments. Future work includes estimating inflow coefficients a priori from geographical information and extending toward broad-area, high-frequency estimation through satellite remote sensing integration.
Suspended Particle content in bottom Sediment per unit volume (SPSS) is a representative indicator for assessing red soil impacts, defined as the mass of soil particles smaller than silt per unit volume of bottom sediment (kg/m³) [1]. Administrative monitoring evaluates sedimentation on a nine-level ranking scale based on SPSS to set priority measures and reduction targets [5]. However, SPSS measurement requires sediment sampling and laboratory analysis at high cost; observations by Okinawa Prefecture are limited to approximately twice per year [1].
The red soil runoff process progresses from rainfall onset through deposition, dispersion, and recovery over roughly one day to one week. As Ohmija [1] demonstrated, SPSS generally peaks after the rainy season and decreases where strong typhoon- or seasonal-wind-driven waves resuspend deposited sediment, exhibiting distinct seasonal variations. Biannual observations cannot resolve such dynamics, necessitating higher temporal resolution estimation.
One approach is physical model-based prediction. Sakai and Nakandakari [2] proposed a concise sediment budget model: terrestrial red soil inflow as a positive factor, wave resuspension and tidal flushing as negative factors, sequentially updated by rainfall, wave, and tidal forcing. Physical models decompose individual processes in an interpretable manner applicable to countermeasure evaluation. However, under sparse observations, unrepresented processes, data constraints, and inter-site environmental differences manifest as estimation errors.
In this study, the model of [2] was applied to SPSS data from observation sites in the FY2022 Red Soil Runoff Prevention Verification Project [3], combined with Japan Meteorological Agency data. Each site was paired with its nearest meteorological station for forcing inputs.
Comparing baseline models (BL1: Persistent; BL2: Linear regression) with the proposed model (A1), BL2 achieved the best simple RMSE, but A1 performed best under Log-RMSE. Given that SPSS spans a wide range following a log-normal distribution, A1's Log-RMSE superiority suggests relatively uniform accuracy across both low and high value ranges.
We then examined performance when parameters set uniformly across all sites in [2] were optimized per site. For the river sediment inflow coefficient A, site-specific optimization improved accuracy, whereas for the wave resuspension coefficient M, accuracy remained virtually unchanged with a uniform value. This reflects that red soil inflow depends strongly on geographical factors (catchment area, land use, river morphology) with large inter-site variability, while wave resuspension is governed by relatively homogeneous mechanics. Accurate characterization of site-specific river inflow properties is thus key to SPSS estimation.
These results demonstrate that site-specific river inflow parameter calibration directly improves physical model-based SPSS estimation, and that broad-area application is feasible with minimal parameter adjustments. Future work includes estimating inflow coefficients a priori from geographical information and extending toward broad-area, high-frequency estimation through satellite remote sensing integration.