This repository contains the Machine Learning implementation for Work Package 2 of the WATER project, funded by the European Space Agency (ESA).
The goal of this package is to transition from physical-constraint-based detection to a data-driven approach using Deep Learning (U-Net) to identify "active" mesoscale eddies.
The WATER project aims to quantify and enrich the mesoscale content extracted from conventional altimetry and experimental SWOT-enriched data.
Active Eddies are defined as coherent sea-level anomaly patterns that possess a significant, co-localized environmental anomaly in Sea Surface Temperature (SST). This approach filters out spurious noise and focuses on physically/biologically active structures.
We implement a multi-task U-Net that simultaneously performs:
- Binary Classification: Detection of the eddy mask.
- Regression: Reconstruction of the co-localized SST spatial anomaly.
To ensure physical consistency, the model is trained using a composite loss function:
- Weighted BCE: For the eddy mask.
- MSE: For the SST anomaly reconstruction.
- Boundary-SST Coupling Loss: A custom term that forces the mask branch to align its borders with high SST gradients.
- Gradient Loss: Ensures sharp edges in the reconstructed temperature fields.
Based on current evaluations (see eval/):
- Nadir-U-Net: Achieves high F1-scores (~0.81) on conventional altimetry.
- SWOT-U-Net: Fine-tuned on SWOT-enriched data, detecting up to 30% more eddies compared to dynamical algorithms when applied to nadir data.
- Explainability: The model leverages the SST anchor to identify high-resolution patterns even from low-resolution altimetry inputs.
src/: Core library (Model, Losses, Data Generators).experiments/: Training scripts for Nadir (full/pre-train) and SWOT (fine-tuning).eval/: Scripts for binary/regression metrics and logit threshold optimization.
git clone https://github.com/YourUsername/WATER-ML-Eddy-Detection.git
cd WATER-ML-Eddy-Detection
pip install -r requirements.txt