222-8 Machine Learning-Based Estimation of Total Suspended Solids Using Field and Hyperspectral Satellite Data

Session: Delta Evolution from Rivers to the Shelf: Past, Present and Future Perspectives for Society (Posters)

Poster Booth No.: 163

Presenting Author:

Onur Karaca

Authors:

Karaca, Onur¹, Khan, Shuhab², Carlson, Brandee³, Wright, Kyle⁴, Aramburu Tinoco, Claudia⁵, Wells-Mourre, Rebekah⁶, Garcia, Sarah⁷

(1) Department of Earth and Atmospheric Science, University of Houston, Houston, TX, USA, (2) Department of Earth and Atmospheric Science, University of Houston, Houston, TX, USA, (3) Department of Earth and Atmospheric Science, University of Houston, Houston, TX, USA, (4) Texas Water Development Board, Texas Water Development Board, Houston, TX, USA, (5) Department of Earth and Atmospheric Science, University of Houston, Houston, TX, USA, (6) Department of Earth and Atmospheric Science, University of Houston, Houston, TX, USA, (7) Department of Earth and Atmospheric Science, Uniersity of Houston, Houston, TX, USA,

Abstract:

Accurate and spatially extensive monitoring of Total Suspended Solids (TSS) is critical for evaluating sediment transport, water clarity, and aquatic ecosystem health in estuarine environments. This study presents a comprehensive remote sensing and machine learning framework for TSS estimation in two dynamic estuarine systems along the Texas coast: Matagorda Bay and Trinity Bay. A total of 110 water samples and subsurface reflectance spectra were collected during monthly field campaigns between August 2024 and July 2025, using an ASD FieldSpec® spectroradiometer (400–900 nm). Reflectance profiles showed strong positive correlations with TSS, especially in the visible range (400–700 nm), enabling robust model development. Five machine learning algorithms—Linear Regression (manual and automated), Random Forest, XGBoost, LightGBM, and CatBoost—were trained on field data. CatBoost achieved the highest accuracy (R² = 0.978, RMSE = 3.015 mg/L), with SHAP analysis revealing red and green spectral bands as the most influential predictors across models. Trained models were applied to PACE, PRISMA and EMIT hyperspectral imagery to produce high-resolution TSS maps for both bays. CatBoost and linear models effectively captured spatial sediment variability, aligning with field observations, whereas the LightGBM model performed less effectively in areas with dynamic sediment conditions. This study shows that combining field spectroscopy, machine learning, and new hyperspectral satellites is a powerful way to monitor TSS in coastal areas. The method can help improve sediment management and protect water quality over time.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-8509

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.