TL;DR
This paper evaluates eight different methods for deriving shallow water bathymetry (depth) of coral reefs using satellite data, combining optical imagery (Sentinel-2) with laser altimetry (ICESat-2). Machine learning models, particularly Multilayer Perceptron (MLP), generally outperformed traditional methods, offering more accurate and detailed depth maps, which is crucial for monitoring these vulnerable ecosystems.
What Problem Does This Paper Address?
Accurate bathymetric maps are essential for understanding and managing marine ecosystems, especially coral reefs, which are vital for biodiversity and coastal protection. However, traditional methods like ship-based sonar or airborne lidar are expensive, time-consuming, and often impractical for remote or politically sensitive areas like the Spratly Islands. Satellite-derived bathymetry (SDB) offers a promising alternative, but choosing the most effective SDB algorithm, especially when fusing active (ICESat-2) and passive (Sentinel-2) sensors, remains a challenge due to a lack of comprehensive comparative studies. This paper aims to fill this gap by rigorously evaluating various SDB approaches in a data-scarce coral reef environment.
Methods
The study employed a multi-algorithm evaluation framework. The core methodology involved:
- Data Preprocessing: Sentinel-2 Level 2A images were processed and synthesized into a single, cloud-free composite image using a median filter on the Google Earth Engine (GEE) platform. This step is crucial for minimizing image-induced uncertainties.
- Active Sensor Data Extraction: ICESat-2 ATL03 data, which provides photon-level elevation information, was meticulously processed. This included:
- Manual Annotation: Using the PhotonLabeler software for accurate seafloor photon identification, addressing noise in the raw data.
- Refraction Correction: Applying a model to account for the bending of laser light at the air-water interface.
- Tidal Correction: Referencing the depth measurements to mean sea level.
- Spatial Aggregation: Stacking the corrected ICESat-2 points into 10m pixels to match the Sentinel-2 imagery resolution.
- SDB Model Implementation: Eight different SDB models were implemented:
- Classical Methods: Lyzenga and Stumpf models, which are established techniques for bathymetry estimation.
- Machine Learning Methods: Six algorithms were tested: Multilayer Perceptron (MLP), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Light Gradient Boosting Machine (LGBM).
- Model Training and Validation: A portion of the ICESat-2 data was used for training the models, while an independent dataset of in-situ sonar measurements was used for validation. Hyperparameter tuning was performed using grid search cross-validation.
- Accuracy Assessment: The performance of each model was evaluated using standard metrics: Coefficient of Determination (R²), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).
Datasets
- Sentinel-2 Imagery: 12 high-quality Sentinel-2 images (four per year from 2019-2021) were used. These images provide multispectral information crucial for optical-based bathymetry.
- ICESat-2 ATL03 Data: This dataset provided precise, geolocated photon data from the ICESat-2 satellite, serving as the primary source for active-sensor bathymetric points.
- In-situ Sonar Data: Measured sonar data from the Odom Hydrotrac II system, collected on 21 January 2018, was used as the ground truth for validating the SDB models. This dataset covered a route length of 12.45 km and provided depth measurements referenced to mean sea level.
Key Insights and Ideas
- Machine Learning Dominance: Machine learning algorithms consistently outperformed classical methods in terms of accuracy and the ability to capture finer bathymetric details. This highlights their superior capacity for handling complex, non-linear relationships between spectral data and water depth.
- MLP as a Top Performer: The Multilayer Perceptron (MLP) model emerged as the most accurate, achieving the lowest RMSE, indicating its strong potential for reliable bathymetric retrieval in optically shallow waters.
- Depth-Dependent Accuracy: All models showed a general decrease in accuracy with increasing water depth. This is an inherent limitation of optical SDB, as light penetration decreases in deeper waters, and also affects the quality of ICESat-2 seafloor returns.
- Classical Models' Limitations: While classical models provided acceptable results, they tended to underestimate depths, particularly in deeper regions (15-20m), and showed less spatial detail compared to ML models.
- Importance of Data Quality: The study emphasizes the critical role of high-quality, cloud-free satellite imagery and accurate active sensor data (ICESat-2) for successful SDB. The median compositing technique was vital for overcoming image noise.
Key Results
- Overall Accuracy: All eight SDB models demonstrated good agreement with measured depths, with R² values exceeding 0.91.
- Machine Learning Performance: Machine learning models achieved a mean R² of 0.943 and an RMSE ranging from 1.034 m to 1.202 m.
- Top Models: MLP achieved the lowest RMSE of 1.034 m, followed closely by KNN (1.070 m).
- Classical Model Performance: Lyzenga and Stumpf models exhibited MAE and RMSE values of less than 0.990 m and 1.386 m, respectively, for the 0-20m range. However, Lyzenga showed a significantly higher RMSE (3.233 m) in the 15-20m depth range compared to other models.
- Depth-Specific Accuracy: In the 15-20m depth range, MLP achieved an RMSE of 1.367 m, while Lyzenga's RMSE was 3.233 m, underscoring the performance gap.
Limitations and Future Work
- Depth Limitations: The accuracy of SDB models generally decreases beyond 20 meters, and the availability and quality of ICESat-2 data can be affected by cloud cover and water turbidity.
- Data Scarcity: While this study benefited from ICESat-2 training data, machine learning models can face limitations in areas with very sparse training data.
- Environmental Heterogeneity: Machine learning models might be sensitive to environmental variations and may require recalibration for different locations with distinct characteristics.
- Future Directions: The authors suggest exploring more advanced machine learning or deep learning techniques for bathymetric retrieval, especially for complex water types and coastal regions. Further research into the feasibility and robustness of these methods is recommended.
Summary updated