Article

Global estimating total suspended solids concentration in inland and coastal waters using deep learning

Details

Citation

Huang S, Li X, Jiang D, Yang Y & Li X (2026) Global estimating total suspended solids concentration in inland and coastal waters using deep learning. Journal of Oceanology and Limnology. https://doi.org/10.1007/s00343-026-6002-y

Abstract
Total suspended solids (TSS) concentration is a fundamental indicator of water turbidity, sediment dynamics, and aquatic ecosystem health in inland and coastal waters. Accurate satellite-based TSS retrieval for those water bodies remains challenging due to their complex bio-optical conditions, high variability, and uncertainties in atmospheric correction. To address these challenges, we develop Total Suspended Solids Estimation Network (TSS-Net), a deep learning framework for global TSS estimation from multispectral remote-sensing reflectance (Rrs). TSS-Net employs a multi-layer feedforward architecture with residual connections to preserve spectral information and enhance model stability and generalization. A three-stage hierarchical training strategy is adopted by integrating radiative transfer simulations, global in-situ measurements, and Sentinel-2 MSI satellite observations. The radiative transfer simulations were used to physically constrain TSS-Net. Specifically, HydroLight-EcoLight 6 simulations are firstly used for pre-training to establish physically consistent Rrs-TSS relationships. The model is subsequently fine-tuned using a global representative in-situ dataset (1 776 samples) to capture real-world optical variability, and further adapted with 1 129 quality-controlled Sentinel-2 satellite matchups to mitigate satellite-specific uncertainties. Extensive evaluations demonstrate that TSS-Net consistently outperforms existing widely-used empirical, semi-analytical, and machine learning models, including the Jiang, MDN, and SOLID approaches. Evaluation using an independent global in-situ dataset, TSS-Net achieves R2 of 0.84, mean absolute percentage error (MAPE) of 49.94%, and near-zero bias. When applied to Sentinel-2 matchups, it maintains strong performance (R2=0.73, MAPE=37.39%), indicating robust generalization across diverse optical environments. Applications to the Upper Chesapeake Bay show that TSS-Net produces physically plausible TSS distribution patterns without region-specific recalibration. Overall, TSS-Net provides a robust, scalable, and generalizable solution for global TSS retrieval, supporting long-term water-quality monitoring.

Keywords
total suspended solid (TSS); color remote sensing; Sentinel-2; deep learning

Journal
Journal of Oceanology and Limnology

StatusEarly Online
FundersNational Natural Science Foundation of China
Publication date online31/10/2026
Date accepted by journal28/04/2026
PublisherSpringer Science and Business Media LLC
ISSN0254-4059
eISSN1993-5005

People (1)

Dr Dalin Jiang

Dr Dalin Jiang

Research Fellow, Biological and Environmental Sciences