Long-Lead ENSO Forecasting with a Seed-Ensembled U-Net

Abstract

We study year-ahead prediction of monthly El Niño-Southern Oscillation (ENSO) sea-surface temperature (SST) anomalies using a four-scale U-Net trained on ERA5 (1940-2024) anomalies defined relative to the 1951-1980 climatology and remapped to a 64 × 128 grid. The model uses a 12-month context and forecasts the next 12 months auto-regressively via a delta-over-persistence head. Training uses per-grid-cell standardization, an area-weighted MSE, and an event-aware sample weight that up-weights months with large Niño-band anomalies. We run controlled ablations over random seeds, batch size, input scalers, model width, spatial loss footprint, regional loss blending, and backbone choice (U-Net, ResNet, U-Net++, Attention U-Net, ConvLSTM, UNet ConvLSTM). Across Niño regions and leads, the U-Net consistently attains the strongest anomaly-correlation skill (ACC); simple seed ensembling further improves performance over the best individual seed. Standardization outperforms Min-Max and IQR-based scaling; a global loss with event weighting dominates tropics- or equatorial-Pacific-restricted losses; and wider channels offer only modest gains over a slim baseline. The resulting configuration—U-Net, StandardScaler, global loss with event weighting, and a small seed ensemble—yields the highest mean ACC while maintaining low global error. We also document observed couplings between SST and ERA5 surface sensible heat flux, low-level divergence, and moisture flux for physical context; these fields are not yet used as inputs but suggest a path toward physics-informed forcings for large events.