Researchers Predict Sea-Surface Debris Movement Around Malta Using AI and Physics Models to Aid Marine Conservation

By Dr Silpaja Chandrasekar, PhDReviewed by Joel ScanlonSep 13 2024

Leveraging AI and physics, this study offers a cutting-edge approach to predicting and managing sea-surface debris, paving the way for more effective cleanup operations and marine ecosystem protection.

Study: Predictive Modelling of Sea Debris around Maltese Coastal Waters. Image Credit: ecstk22 / Shutterstock

In a paper published in the journal Oceans, researchers proposed a machine learning (ML)--based system combined with a physics-based model to predict and visualize sea-surface debris movement around Malta. Using historical sea-surface current (SSC) velocities, two models forecasted conditions for 24 hours, feeding predictions into a Lagrangian model to simulate debris movement.

A comparative evaluation determined the best-performing model, offering a tailored solution to improve cleanup and conservation efforts. This approach enhanced the prediction of SSC and aided in debris management around the island.

Background

Past work highlighted the significant environmental threat posed by sea-surface debris around Malta. This debris, which is composed mainly of plastics, harms marine life and disrupts ecosystems.

Predictive models for marine debris movement have incorporated numerical simulations and deep learning techniques, improving accuracy in forecasting debris dispersal. However, challenges remain in adapting these models to local conditions and accurately predicting complex real-time debris movements.

Predicting Marine Debris

In this study, the approach began with preprocessing the SSC dataset to ensure data accuracy before applying ML models capable of predicting SSC velocities. The SSC dataset from the University of Malta contained hourly measurements of west-east and north-south velocities collected over four years using high-frequency radar systems.

These radar systems around the Maltese islands and southern Sicily provided a geographic grid of 180 data points processed in the network common data form (NetCDF) format. The data underwent preprocessing to merge the temporal and spatial dimensions and address missing values caused by radar limitations near coastal areas.

Although various interpolation techniques were tested, it was concluded that the original dataset should be preserved with NaN values to maintain the realism of simulations.

The Lagrangian model was used to simulate the movement of marine debris, relying on the predicted SSC data to track particle trajectories. Python toolkits like OceanParcels facilitated this process, allowing for custom kernels and behaviors such as particle reflection and boundary interactions.

Particles representing debris were initialized near a specific location in a clustered formation, and their movements were simulated over seven days.

A probabilistic model was introduced to enhance realism, reflecting particles off land with an 85% chance of returning to the sea and a 15% chance of beaching. While wind data were not included, future simulations may incorporate wind and wave data to increase accuracy.

A pipeline was developed to predict SSC velocities using long short-term memory (LSTM) and gated recurrent unit (GRU) models, chosen for their effectiveness in time-series data processing.

The evaluation of these models involved error metrics such as mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). Initial attempts to train models using a year's worth of data yielded suboptimal results, leading to an expanded dataset covering February 2020 to August 2023.

Geospatial filtering focused on 37 data points in a specific area of Malta's northern coast, addressing issues like radar interference near the beach.

The team built separate models for the u and v components of velocity to improve prediction accuracy, and early stopping was applied to prevent overfitting.

A rolling forecasting method was used, with 72 hours of data inputted to predict the next 24 hours. The model architecture, including layers of LSTM/GRU, dropout, and dense layers, was tuned for optimal performance using TensorFlow, with error metrics such as mean absolute error (MAE), mean squared error (MSE), and RMSE used to assess accuracy.

The pipeline integrated predictions with a physics-based Lagrangian model to simulate debris dispersion over 24 hours. Using forecasts from both LSTM and GRU models, simulations were visualized to compare debris movement on two test dates: August 4 and November 4, 2023.

Although drifter data was unavailable for direct validation, geospatial heat maps, and centroid analysis were used to evaluate the spatial accuracy of the model predictions.

The results indicated that LSTM models generally provided more consistent and reliable forecasts, especially in regions with denser data points. The spatial spread, centroid, and skewness of the models' outputs were also analyzed, revealing differences in the consistency and directional tendencies of the LSTM and GRU models.

Model Performance Evaluation

The primary goal was to determine which model, LSTM or GRU, performed better in predicting SSC velocities and to evaluate the real-world accuracy of the Lagrangian simulations based on these predictions. Analysis of error metrics—MAE, MSE, and RMSE—revealed that LSTM generally outperformed GRU.

In Test 1 (August 4), LSTM showed slightly lower MAEs and RMSEs for the u component, indicating better accuracy and consistency, although GRU had a marginally lower MSE. For the v component, both models were similar in performance, but LSTM had a lower standard deviation and IQR, indicating more reliable predictions.

In Test 2 (November 4), both models had high error metrics for the u component, but LSTM again demonstrated better performance with lower MAEs, MSEs, and RMSEs. For the v component, error metrics were higher across both models, with GRU showing worse results. Overall, LSTM's more consistent performance, even under varying seasonal conditions, suggests it is the more reliable model.

Conclusion

To sum up, this study integrated ML models with a physics-based Lagrangian framework to predict sea-surface debris movement around Malta. It found that the LSTM model outperformed GRU in accuracy and reliability.

Despite challenges like missing data and limited validation, the LSTM model proved effective and consistent across different conditions.

The study also explored future applications of this predictive framework, such as modeling jellyfish dispersion, maritime pollution management, and optimizing response strategies for environmental cleanup operations.

Future work could enhance predictions with additional parameters and explore broader applications, including jellyfish dispersion and maritime pollution management.