Comparative Analysis of Prediction Accuracy for Drifting Buoy Data Using CNN (Conv1D) and GRU Deep Learning Models with Varying Data Volumes

Abstract

Drifting buoy data plays a vital role in climate and oceanographic research, offering critical insights into ocean surface dynamics, currents, and weather patterns. Accurate trajectory prediction of drifting buoys improves maritime weather forecasting, supports climate change research, and aids in search and rescue operations at sea. This research explores the application of two deep learning models, CNN (Conv1D) and GRU (1D), to predict buoy trajectories. The study utilizes the timestamped geographical coordinate datasets, which were processed and divided into training and testing sets. Both models were optimized using the Adam algorithm and Huber loss function, with hidden layer filter configurations of 64, 128, and 256. Model performance was evaluated using MSE, RMSE, MAE, R², Cohen’s Kappa, and F1-score. Results indicate that CNN (Conv1D) consistently outperforms GRU, particularly with 256 filters, achieving significantly lower RMSE and MAE values, demonstrating higher predictive accuracy. While GRU exhibited performance fluctuations across different filter configurations, CNN (Conv1D) maintained stable accuracy across varying dataset conditions. Notably, CNN (Conv1D) achieved at least 50% greater accuracy than GRU while preserving a near-perfect correlation to input data. The study highlights the critical role of high-resolution data in enhancing prediction reliability, as lower-resolution or highly variable datasets negatively impact model performance. Additionally, it underscores the importance of effective preprocessing techniques for handling missing data to ensure robust predictions. This research advances deep learning applications in marine studies by optimizing trajectory forecasting models. Future work should explore hybrid approaches integrating Conv1D with other architectures or leveraging transformer models to enhance long-term prediction accuracy. These findings provide a reliable framework for oceanographic research, maritime navigation, and environmental monitoring in dynamic marine conditions.