Feature-Gradient Enhanced Transformer for Accurate dissolved oxygen saturation Forecasting in Marine Environments

Weiyan Tan^*Bing GengXiuGuang Bai

• Guangdong service center for veterans, Guangzhou, China

With the increasing severity of marine environmental issues, dissolved oxygen saturation (O2%) forecasting is not only essential for water quality monitoring and ecological risk assessment, but also for deepening our understanding of how physical and biogeochemical processes jointly shape coastal oxygen dynamics. Using multi-year in situ records from five representative NOAA ocean observation stations spanning distinct hydrographic regimes, this study develops an improved time series forecasting framework based on the Transformer architecture and uses it as a lens to analyze the structure and predictability of O2% variability. A Feature Pyramid Space Transformation (FPST) is incorporated into the encoder to decompose the observed O2% time series into multiple temporal scales, enabling the identification of station-dependent contributions from long-term trends, synoptic variability, and short-term fluctuations. In the decoder, a Gradient Attention Mechanism (GAM) explicitly leverages temporal gradients to highlight sharp transitions and turning points in the observational record, thereby revealing how rapid changes and extreme episodes affect the local predictability of O2%. Experiments on the five buoy datasets show that the proposed framework achieves consistently improved forecasting performance over a range of baseline methods; for example, averaged across all stations, Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) are reduced while the coefficient of determination (R2) is increased, indicating a better fit between predicted and observed time series. Further ablation and input-subset experiments demonstrate that FPST and GAM provide complementary benefits and elucidate the relative importance of depth, temperature, and local oxygen concentration as drivers of O2% dynamics at different sites. Overall, the study offers both a robust forecasting tool and an observation-based characterization of the multi-scale structure and event-driven behaviour of dissolved oxygen saturation in coastal ocean environments.