Cross-Regime Transfer Learning for Dekadal Rainfall Prediction in Indonesia Using a Transformer Encoder

Abstract

. Rainfall prediction across regions with different rainfall regimes remains challenging because a model trained in one climatic domain can fail when applied to another domain. This study evaluates transfer learning for dekadal rainfall prediction using a Transformer encoder trained on 10 rainfall sites in West Java and adapted to three target regions representing different rainfall patterns: Kupang (monsoonal), Padang (equatorial), and Ambon (local). Three transfer strategies were evaluated: zero-shot inference, head-only fine-tuning, and staged unfreezing. The source model used 22 meteorological and climate-index features, site embedding, a 36-step input window, and H1-H6 forecasting horizons. Zero-shot transfer produced negative R² on all targets, indicating substantial domain shift. Staged unfreezing consistently improved the zero-shot model, reducing RMSE by 49.83% in Kupang, 9.13% in Padang, and 24.76% in Ambon. On the raw millimeter scale, the final transferred model obtained RMSE values of 69.30 mm, 168.36 mm, and 92.91 mm for Kupang, Padang, and Ambon, respectively. Compared with simple baselines, the transferred model outperformed persistence on all targets but outperformed climatology only in Ambon. These results indicate that transfer learning is useful for adapting deep rainfall models to new regions, but reliable deployment still requires target-domain fine-tuning, raw-scale evaluation, and comparison against simple climatological baselines.