Two-Stage TSS-Integrated 3D-CNN Framework for Satellite-Derived Bathymetry in Turbid Coastal Waters

Abstract

The accuracy of satellite-derived bathymetry (SDB) is known to decline in turbid coastal waters, where elevated Total Suspended Solids (TSS) concentrations alter the optical properties of the water column and weaken the spectral–depth relationship on which conventional SDB algorithms rely. To address this limitation, this study develops a two-stage 3D-CNN model that integrates TSS as an additional parameter derived from Sentinel-2 imagery. The framework was developed and evaluated using 1,486 bathymetric points and 26 TSS samples collected synchronously with Sentinel-2 acquisition on 13 August 2025 in the coastal waters of Pekalongan, Central Java, Indonesia, across four patch sizes (9×9, 12×12, 15×15, and 16×16 pixels) under a 5-fold cross-validation scheme. The first model (CNN-TSS) achieved near-perfect retrieval accuracy (R² = 0.999–1.000; RMSE = 0.135–0.194 mg/L), while the second model (CNN-SDB) (9×9, without TSS) achieved R² = 0.931 (RMSE = 0.276 m). Integrating predicted TSS improved accuracy only at the 15×15 patch size (R² = 0.934, RMSE = 0.271 m). All CNN configurations substantially outperformed conventional SDB algorithms, reducing RMSE by up to 59.97%. This study offers the first systematic patch-size sensitivity analysis for TSS-integrated CNN-based bathymetry, elucidating when auxiliary water-quality information enhances or degrades deep-learning depth retrieval.