Answers are grounded in this paper and cite their sources.
: This study presents a neural network model for estimating shallow-water depths using Sentinel-2 multispectral data and bathymetric measurements. Unlike traditional approaches of satellite-derived bathymetry (SDB), the model incorporates date, time, and location of the dataset, capturing both spatial and temporal variability. The input features include spectral bands (B1, B2, B3, B4), geographic coordinates, and date and time of image acquisition, with depth predictions focused on the 0.0 to 10.0 m range. The trained model demonstrated strong predictive accuracy, achieving a test mean absolute error (MAE) of 0.47 m. However, the external model validation (multi-beam bathymetry) highlighted regional variations in MAE, particularly between open waters and estuarine environments. The greatest residual variability was observed in highly turbid estuarine areas, emphasizing the need for depth range filtering to maintain prediction reliability. The model achieved a maximum correlation coefficient of 0.95 in open waters, demonstrating strong agreement between predicted and observed depths. External validation, which focused on shallower regions of open waters, was conducted using both single-beam echo sounder and ICESat-2 depth measurements. This validation further confirmed the model’s accuracy, with residuals consistently within ±0.5 m. These findings highlight the potential of deep learning for satellite-derived bathymetry in dynamic and turbid environments. The model’s ability to adapt to regional and temporal variability makes it a valuable tool for coastal monitoring and remote sensing applications, supporting more accurate and consistent bathymetric estimations over time.
No takes yet. Share an insight, caveat, or question.
Santos et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: