Blending physical and artificial intelligence models to improve satellite-derived bathymetry mapping
Bathymetry is a fundamental source of information for understanding the marine environment and serves as the primary basis for initiating projects related on marine cartography. Satellite-Derived Bathymetry (SDB) offers a promising alternative to address the lack of current bathymetric data in shall...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125003371 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Bathymetry is a fundamental source of information for understanding the marine environment and serves as the primary basis for initiating projects related on marine cartography. Satellite-Derived Bathymetry (SDB) offers a promising alternative to address the lack of current bathymetric data in shallow areas, particularly when utilizing data from satellite platforms such as Sentinel-2, along with advanced Machine Learning and Deep Learning techniques capable of enhancing bathymetric estimation compared to conventional methods. In this study, we evaluated the performance of two different algorithms for estimating SDB in two areas of the Western Mediterranean: a physics-driven model and an Artificial Neural Network (ANN). We assessed the ability of these methods to predict bathymetries over successive years subsequent to algorithm calibration, as well as their capacity to estimate depths of other areas not included in model calibration, thereby evaluating temporal and spatial independence, respectively. In situ depth measurements collected by echo sounders in the study areas were used to train and test the algorithms. Performance metrics, including the coefficient of determination (R2) and Root Mean Square Error (RMSE), consistently yielded results with R2 ≥ 0.8 and RMSE ≤1.5 m across all sections, demonstrating a strong goodness of fit even under scenarios designed to assess temporal and spatial independence. Furthermore, both models showed good performance in estimating both shallow (<10 m) and deeper (>10 m) waters, achieving Median Absolute Error (MedAE) ≤ 0.82 m and Mean Bias Error (MBE) ≤ 0.82 m for shallow waters, and MedAE ≤2.15 m and MBE ≤ 2.33 m even when the depth range extends up to 20 m. The ANN demonstrated a slight improvement in precision, particularly advantageous when a sufficient amount of in situ data is available for algorithm training and validation. Conversely, the physics-driven method, despite being slightly less precise than ANN, offered advantages, especially in study areas with limited data availability. |
|---|---|
| ISSN: | 1574-9541 |