Huidong (Warren) Jin: Machine Learning for High-Resolution Long-Range Daily Rainfall Forecasts from ACCESS-S2
Machine Learning for High-Resolution Long-Range Daily Rainfall Forecasts from ACCESS-S2
Authors: Huidong (Warren) Jin¹, Rui Wang², Yaozhong Liu², Xinni Song², Quanxi Shao³, Ming Li³
Affiliations: ¹ CSIRO Data61, Canberra ACT; ²Australian National University, Canberra ACT; ³ CSIRO Data61, Kensington WA
Presenting author: Warren Jin
- Introduction and Scientific Context
Kilometre-scale daily rainfall forecasts for the coming weeks to months are critical for agriculture, water management, and emergency services across Australia. The Bureau of Meteorology’s ACCESS S2 seasonal prediction system provides ensemble rainfall forecasts at ~60 km resolution. However, its raw outputs are overly smooth, poorly calibrated, and systematically underestimate extreme daily rainfall, limiting their value for local decision-making. - Methodology
We develop two AI-based post-processing models to produce spatially realistic and probabilistically skilful 5 km daily rainfall forecasts at lead times of 0–41 days, trained on 1990–2005 hindcasts and evaluated against AWAP observations.
DESRGAN (Downscaling with Enhanced Super-Resolution GAN) uses cascaded Residual-in-Residual Dense Blocks and a U-Net discriminator to achieve 12× spatial upscaling. It outperforms quantile mapping and climatology on CRPS and MAE but produces deterministic outputs and reduces ensemble reliability.
PRGAN (Probabilistic Rainfall GAN) addresses these limitations by generating Bernoulli–Gamma distribution parameters at each grid point, enabling direct exceedance probability estimation. A two-stage training strategy—adversarial pretraining followed by negative log-likelihood fine-tuning—stabilises learning for zero-inflated rainfall. - Results
Across El Niño (2006), La Niña (2007), and neutral (2018) years, PRGAN consistently outperforms baselines on MAE, CRPS, and Brier scores at extreme thresholds (95th, 99th, 99.5th percentiles), while substantially improving ensemble reliability. - Audience
Machine learning researchers, climate scientists using ACCESS-S2, and practitioners in agriculture, water management, and hazard assessment. - Keywords
Seasonal climate forecasting; generative adversarial networks; probabilistic downscaling; extreme rainfall; ACCESS-S2