| Abstract | Marine stressors such as temperature, salinity, turbidity, and chlorophyll strongly influence the health of blue carbon ecosystems (BCEs) and play a critical role in global climate regulation. Accurate long-term forecasting of these stressors enables baseline construction, anomaly detection, and restoration planning. In this study, we present forecasting using spatio-temporal transformer (FUSTT), a deep learning model-derived from making efficient additions to the original transformer architecture, specifically for multivariate long-term forecasting of marine parameters. FUSTT integrates a dual-encoder design, dimensional flipping layer, and flagged dual stage attention, enabling the model to capture complex temporal and cross-variable dependencies. Using real-world datasets collected by ocean networks canada at four representative Salish Sea locations (Discovery Passage, Burrard Inlet, Baynes Sound, and Boundary Pass), FUSTT consistently outperformed baseline models. It achieved 15–33% lower error than traditional transformer architectures and 30–40% higher accuracy compared to domain-specific models such as LSTM and CNN-LSTM, while also delivering faster training with minimal tradeoffs. These results demonstrate FUSTT’s capability to provide highly accurate and computationally efficient predictions of marine stressors at operational ecological monitoring sites, offering a practical foundation for advancing BCE management and climate resilience. |
|---|
