DriftNet: target-area differential attention mechanism for marine drifter trajectory prediction

Recent years have seen the significant potential of Artificial Intelligence (AI) techniques in monitoring and managing ocean ecosystem. Among these, deep learning (DL) solution demonstrated considerable performance in forecasting drifter trajectory, which provide critical scientific support in a variety of ways. However, most existing DL methods still suffer from stability and accuracy problems, due mainly to the fact that they rely primarily on historical memory of tracking target, and that attention-introduced bias further limits their applicability. To address these challenges, this study proposes an innovative drifter trajectory prediction framework (DriftNet) based on newly designed 1) Target-Area Differential Attention (TADA) mechanism and, 2) Direction-Distance Loss function (DDL). For TADA, we constructed a tempo-spatial structure by combining the initial target position and surrounding environment features, then we introduced a Differential Transformer to the integrated features. DDL is an improved loss function to jointly optimize directional alignment and spatial displacement in drift trajectory prediction. Utilizing this model, a case study was conducted in the Taiwan Strait, and the results showed that DrifterNet improved drifter prediction accuracy by over 50% compared to the Lag-fjhyj operational model. This achievement indicates that DriftNet is not only able to better capture the features of target and surrounding area, but better integrate the interaction of timestamp data and spatial data, which further deepens our understanding of the interaction between time-domain factors and spatial-domain factors in maritime environment. This study presents a novel and effective solution for marine drift prediction, offering both theoretical insights and practical value for real-world ocean monitoring and emergency response scenarios.