Zeus D. Elderfield, Seth S. Demeterio, Isabel Joy Adriatico, Gerard Durano Ompad, Christian V. Maderazo
Marine heatwaves (MHWs) - prolonged sea surface temperature (SST) departures above seasonal thresholds - are increasing in the Philippine Sea, threatening fisheries, reefs, and coastal livelihoods. The researchers assemble a daily, area-averaged 1995-2024 record that merges satellite SST, large-scale climate indices (ENSO, PDO, IOD, MJO), and ERA5 atmospheric fields, interpolate short gaps, standardize variables, and use 90-day lookbacks with a one-day (lead-1) forecast horizon. Recursive feature elimination identifies five informative drivers: SST anomaly, Indian Ocean Dipole, surface solar radiation, 2-meter air temperature, and 10-meter meridional wind. A univariate N-BEATS baseline (SST only) achieves lead-1 MAE 0.195 °C, RMSE 0.247 °C, R2 0. 9 5 9, and MHW F1 0.937 on the held-out test block. A two-stream N-BEATSX that encodes exogenous histories reduces lead-1 errors to MAE 0.087 °C and RMSE 0.108 °C (R2 0.992) while raising MHW F1 to 0.971. Kernel SHAP attribution shows the retained drivers-winds, air temperature, solar radiation, and SST anomaly-jointly explain roughly half of the model's predictive variance alongside the autoregressive SST channel. These results demonstrate that integrating carefully selected exogenous variables into an interpretable deep-learning framework substantially improves near-term MHW predictability and highlights the physical drivers modulating Philippine Sea heat extremes. © 2025 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
