5:15 PM - 7:15 PM
[ACG48-P07] Satellite-based estimate of river suspended sediment without calibration requirement
Keywords:Suspended Sediment, Satellite Observation
Understanding the sediment discharge from global rivers into the ocean is crucial for quantifying sediment budgets between terrestrial and coastal regions and elucidating nutrient dynamics associated with sediments. Furthermore, nutrients transported by rivers serve as a major source for marine organisms, which play a vital role in carbon dioxide absorption, making this process essential for climate change predictions. However, rivers with continuous suspended sediment concentration (SSC) monitoring account for less than 10% of the global total, and comprehensive understanding of sediment dynamics at the global scale remains insufficient. Therefore, it is necessary to develop SSC estimation methods that can be applied to unmonitored basins using satellite observations capable of global monitoring.
Previous studies estimating inland water SSC on a global scale using satellite reflectance can be broadly categorized into empirical models, machine learning approaches, and physics-based methods. Empirical and machine learning methods establish relationships between water surface reflectance spectra observed by satellites and in situ SSC measurements; however, these methods are based on limited site-specific observations and lack a foundation for application to unmonitored regions. Physics-based methods primarily utilize data from lakes and coastal areas, where temporal and spatial variations in sediment properties such as particle size distribution and mineral composition are relatively small. In these methods, parameters representing sediment properties are either fixed or classified into a limited number of categories, which poses a challenge for accurate estimation in rivers, where sediment properties exhibit significant spatiotemporal variability.
This study aims to achieve the following objectives: (1) development of an SSC estimation algorithm that considers sediment properties, (2) validation of the algorithm using satellite observations and in situ measurement data.
(1) Algorithm Development: A forward model representing the relationship between SSC and reflectance was constructed considering sediment properties. The SSC was then inversely estimated by optimizing the concentration that minimizes the error between observed and modeled reflectance. The core of the forward model is a bio-optical model (Lee et al., 2011) that computes reflectance at each wavelength based on absorption and backscattering coefficients proportional to the concentrations of suspended sediment, water, colored dissolved organic matter (CDOM), and phytoplankton. To enable SSC retrieval, the following enhancements were implemented: (a) Removal of non-sediment influences: CDOM influence was ignored by utilizing wavelengths beyond 665 nm, phytoplankton influence was estimated using an empirical model, and water absorption was represented as a function of water temperature. (b) Constraint of the sediment backscattering coefficient using particle size distribution: The backscattering coefficient was formulated as a function of particle size distribution and wavelength by integrating the backscattering of single spherical particles computed using Mie scattering theory. The particle size distribution was modeled using the 2C-model (Peng and Effer, 2007), which can represent particles smaller than 1 μm. (c) Estimation of sediment absorption coefficient: The sediment absorption coefficient was assumed to remain constant in the near-infrared region. A newly discovered relationship between surface reflectance at 783 nm and 865 nm, which remains site-invariant, was found to be reproducible by maintaining a constant ratio of backscattering to absorption coefficients, allowing for the estimation of sediment absorption coefficients.
(2) Validation: Using Sentinel-2 MSI, SSC was estimated for 635 data points collected from 57 sites across the United States between 2019 and 2023, covering a concentration range of 1.0 mg/L to 2894 mg/L. The algorithm achieved a mean absolute error (MAE) of 138 mg/L, root mean square error (RMSE) of 303 mg/L, and mean absolute percentage error (MAPE) of 94%. Data points within one-third to three times the observed values accounted for 70% of the total, while 191 data points exhibited errors exceeding threefold, with 127 of them located in the Missouri River. A systematic underestimation was observed in the Missouri River during shallow water conditions, suggesting that incorporating bottom reflectance effects could improve estimation accuracy.
Previous studies estimating inland water SSC on a global scale using satellite reflectance can be broadly categorized into empirical models, machine learning approaches, and physics-based methods. Empirical and machine learning methods establish relationships between water surface reflectance spectra observed by satellites and in situ SSC measurements; however, these methods are based on limited site-specific observations and lack a foundation for application to unmonitored regions. Physics-based methods primarily utilize data from lakes and coastal areas, where temporal and spatial variations in sediment properties such as particle size distribution and mineral composition are relatively small. In these methods, parameters representing sediment properties are either fixed or classified into a limited number of categories, which poses a challenge for accurate estimation in rivers, where sediment properties exhibit significant spatiotemporal variability.
This study aims to achieve the following objectives: (1) development of an SSC estimation algorithm that considers sediment properties, (2) validation of the algorithm using satellite observations and in situ measurement data.
(1) Algorithm Development: A forward model representing the relationship between SSC and reflectance was constructed considering sediment properties. The SSC was then inversely estimated by optimizing the concentration that minimizes the error between observed and modeled reflectance. The core of the forward model is a bio-optical model (Lee et al., 2011) that computes reflectance at each wavelength based on absorption and backscattering coefficients proportional to the concentrations of suspended sediment, water, colored dissolved organic matter (CDOM), and phytoplankton. To enable SSC retrieval, the following enhancements were implemented: (a) Removal of non-sediment influences: CDOM influence was ignored by utilizing wavelengths beyond 665 nm, phytoplankton influence was estimated using an empirical model, and water absorption was represented as a function of water temperature. (b) Constraint of the sediment backscattering coefficient using particle size distribution: The backscattering coefficient was formulated as a function of particle size distribution and wavelength by integrating the backscattering of single spherical particles computed using Mie scattering theory. The particle size distribution was modeled using the 2C-model (Peng and Effer, 2007), which can represent particles smaller than 1 μm. (c) Estimation of sediment absorption coefficient: The sediment absorption coefficient was assumed to remain constant in the near-infrared region. A newly discovered relationship between surface reflectance at 783 nm and 865 nm, which remains site-invariant, was found to be reproducible by maintaining a constant ratio of backscattering to absorption coefficients, allowing for the estimation of sediment absorption coefficients.
(2) Validation: Using Sentinel-2 MSI, SSC was estimated for 635 data points collected from 57 sites across the United States between 2019 and 2023, covering a concentration range of 1.0 mg/L to 2894 mg/L. The algorithm achieved a mean absolute error (MAE) of 138 mg/L, root mean square error (RMSE) of 303 mg/L, and mean absolute percentage error (MAPE) of 94%. Data points within one-third to three times the observed values accounted for 70% of the total, while 191 data points exhibited errors exceeding threefold, with 127 of them located in the Missouri River. A systematic underestimation was observed in the Missouri River during shallow water conditions, suggesting that incorporating bottom reflectance effects could improve estimation accuracy.