POSTER SESSION 2
10:30–12:30, Tuesday, September 15
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
ABSTRACT 1166 | POSTER 250
A HYBRID PROCESSING CHAIN FOR SUSPENDED PARTICULATE MATTER ESTIMATION ACROSS EXTREME TURBIDITY GRADIENTS IN THE SEINE ESTUARY
Monitoring suspended particulate matter (SPM) is challenged by extreme concentration gradients (1 to 2000 mg/L). Conventional semi-analytical algorithms are strongly influenced by sediment composition and often fail at extremes: they saturate radiometrically in very turbid waters and are biased by organic absorption in clearer systems. To overcome these limitations, we propose a new framework coupling physical constraints with machine learning. It relies on three harmonized in situ datasets (Seine SYNAPSES, GLORIA archive, and LOG dataset), all quality-controlled and screened with the Quality Water Index Polynomial to remove aberrant satellite remote sensing reflectances. A machine learning model first predicts particulate organic carbon (POC) to compute the POC to SPM ratio, which classifies waters into mineral, mixed, or organic optical types. For each type, SPM is retrieved applying a dedicated semi-analytical equation derived from explainable artificial intelligence diagnostics, mitigating saturation and class-dependent biases. A conditional residual correction is activated only when errors exceed natural bio-optical variability, preventing overfitting. Advanced explainability metrics were exploited to diagnose the physical drivers and ensure radiometric consistency. Following a 70-15-15 data split for training and validation, the framework was evaluated globally and locally on an independent in situ Seine Estuary test dataset (N=52, representing 15 percent of the data), achieving robust performance (log-transformed R2 over 0.83, median errors below 37 percent). Applied to high-resolution Sentinel-2 imagery (10 m), this chain reveals watershed dynamics from upstream to downstream, tracking flood events, wave impacts, and fine-scale structures like water mixing and submerged dikes across all turbidity gradients.
Loic Cabrel Youmbi Tchaewo*, Université du Littoral Côte d’Opale – Laboratoire d’Océanologie et de Géosciences – LOG, [email protected], https://orcid.org/0009-0008-2980-4535
Charles Verpoorter, ULCO/LOG (Université du Littoral Côte d’Opale / Laboratoire d’Océanologie et Géosciences), [email protected], https://orcid.org/0000-0002-5134-4124
Romaric Verney, Ifremer Centre de Bretagne, France, [email protected], https://orcid.org/0000-0001-5650-1726
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
