Bathymetry estimation for coastal regions using self-attention.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Coastal Management

Background:

• Accurate bathymetric mapping is crucial for coastal development and management.
• Conventional bathymetric measurements are costly and constrained.
• Satellite-Derived Bathymetry (SDB) offers a cost-effective alternative but existing models are region-specific.

Purpose of the Study:

• To develop a deep-learning-based SDB model (TransBathy) for accurate bathymetric mapping in diverse, unknown coastal terrains.
• To address the limitations of existing SDB models requiring prior regional knowledge.
• To provide a robust and effective solution for global coastal bathymetry.

Main Methods:

• Developed TransBathy, a deep learning model incorporating an attention mechanism to assimilate deep and spatial features.
• Collected and utilized a large-scale, diverse bathymetric dataset covering multiple shallow coastal regions worldwide.
• Trained the model end-to-end and validated its performance on pre-seen and unseen coastal regions.

Main Results:

• TransBathy achieved good performance on pre-seen regions (RMSE [Formula: see text] m, R² [Formula: see text]).
• Demonstrated effectiveness in challenging unseen regions: RMSE [Formula: see text] m, R² [Formula: see text] (steep slope) and RMSE [Formula: see text] m, R² [Formula: see text] (rugged terrain).
• Outperformed a baseline SDB method in unseen steep slope regions, reducing RMSE by 0.978m and improving R² by 0.187.

Conclusions:

• TransBathy offers a robust and effective solution for satellite-derived bathymetry in diverse and unknown coastal environments.
• The model's ability to generalize across different terrains highlights its potential for global coastal mapping applications.
• The publicly available dataset, code, and weights facilitate further research and application of this advanced SDB technique.