Forecasting Marine Debris Trajectory Based on Lagrangian Particle Tracking Model using Hybrid Deep Learning

Marine debris poses a persistent threat to marine ecosystems and coastal livelihoods, particularly in Indonesia where floating waste disrupts seaweed farming. Traditional Lagrangian Particle Tracking Models (LPTMs) rely on physics-based inputs and lack adaptability to dynamic conditions. This study proposes a deep learning-based approach to forecast marine debris trajectories using historical oceanographic data from Hybrid Coordinate Ocean Model (HYCOM) and weather data from Fifth Generation ECMWF Reanalysis (ERA5). We compare three models Long Short Term Memory (LSTM), Convolutional Long Short Term Memory (ConvLSTM), and Convolutional Neural Network combined with Long Short Term Memory (CNN+LSTM) trained to predict future debris locations from 10 river mouths in South Sulawesi. The CNN+LSTM model outperforms other architectures in short-term forecasting, achieving the lowest mean absolute error (MAE) of 0.5238° and root mean squared error (RMSE) of 0.5711° at step 1, and the lowest mean movement distance difference (MMDD) of 0.0072 km at step 5 during wet season testing. Results demonstrate CNN+LSTM's ability to generalize across seasonal conditions and maintain spatial accuracy. Unlike previous studies, this framework integrates current and weather data for adaptive debris forecasting without requiring predefined physical scenarios. The proposed method offers a practical solution for marine monitoring and coastal risk mitigation.