Spatiotemporal data analysis using deep learning models: A case study with drifting buoy data

Abstract

Building or predicting the trajectory of drifting objects is significant in maritime studies and search and rescue operations. The trajectory of a drifting object can be determined using traditional tools based on marine dynamic models or through artificial intelligence models. Using drifting buoy data collected between December 19 and December 28, 2003, the research team employed a CNN (Conv1D) model for analysis. The results indicate that when using the Adam optimizer, the Huber loss function, and 256 filters in the hidden layer, the model performance parameters were RMSE = 0.04004, MAE = 0.032304 degrees, and R² = 98%. When applying the SGD optimizer and the mean squared error (MSE) loss function, both RMSE and MAE values decreased by up to four times compared to the previous configuration, while the R² value reached 99.9% with 64 filters in the hidden layer. Increasing the number of filters to 128 improved the CNN (Conv1D) model performance by approximately 20%, with RMSE = 0.007863 degrees and MAE = 0.006653 degrees. The R² value approached 100%, indicating that the model is highly suitable for predicting drifting buoy trajectories. Increasing the number of filters from 128 to 256 did not further improve performance, suggesting that 128 filters represent the optimal configuration. However, the RMSE value remains relatively large (0.87 km), possibly due to the limited input dataset. Future studies should consider larger datasets to enhance prediction accuracy.