Bathymetrix-AI
An advanced Machine Learning pipeline for Satellite-Derived Bathymetry (SDB). Features ICESat-2 integration, In-Situ Data filtering, and Spatial Residual Stacking.
Bathymetrix-AI is a specialized QGIS research toolkit designed to derive high-precision bathymetry from corrected multispectral satellite imagery (e.g., Sentinel-2 L2A). It systematically integrates physics-based corrections with data-driven Machine Learning to overcome traditional SDB limitations.
Core Workflow (The 4-Phase System):
1. Automated Pre-processing: Sun-glint removal (Hedley), new Advanced Water Masking (3-Indices), physics-based feature generation (Log-Ratios), and a Deep Water Filter fully customized for ML algorithms.
2. Robust Altimetry Filtering: Uses different algorithms to clean ICESat-2 (ATL24) data by identifying high-confidence inliers and removing environmental noise.
3. Global Auto-ML Modeling: Competitive benchmarking of 11 ML algorithms with Randomized Hyperparameter Optimization to find the optimal global depth function.
4. Spatial Residual Stacking: Enhances accuracy by analyzing prediction residuals and re-training the model with a Stacked Error Surface to correct local biases.
Key References:
Stumpf et al. (2003): Log-Ratio Algorithm for SDB inversion.
Hedley et al. (2005): Physics-based sun-glint correction.
Fischler & Bolles (1981): RANSAC algorithm for ICESat-2 data filtering.
Zhang et al., (2021): LS Variance Fit, or Huber Variance Fit for ICESat-2 data filtering.
Alevizos (2020): Residual analysis and spatial refinement in shallow waters.
Bergstra & Bengio (2012): Randomized search for hyperparameter optimization.
Parrish et al. (2025): Analysis and assessment of global ICESat-2 bathymetry.
Developed for scientific research and hydrographic applications. Creating the Codes and documentation optimized using Google Gemini AI.