Downscaling long lead time daily rainfall ensemble forecasts through deep learning

Abstract Skilful and localised daily weather forecasts for upcoming seasons are desired by climate-sensitive sectors. Various General circulation models routinely provide such long lead time ensemble forecasts, also known as seasonal climate (SCF), but require downscaling techniques to enhance their skills from historical observations. Traditional techniques, like quantile mapping (QM), learn empirical relationships pre-engineered predictors. Deep-learning-based automatically generate select predictors almost all of them focus on simplified situations where low-resolution images match well with high-resolution ones, which is not the case in forecasts. To downscale rainfall we take a two-step procedure. We first choose suitable deep learning model, very super-resolution (VDSR), several outstanding candidates, based an forecast skill metric, continuous ranked probability score (CRPS). Secondly, via incorporating other variables extra input, develop finalise statistical (VDSD) model CRPS. Both VDSR VDSD tested 60 km Australian Community Climate Earth-System Simulator Seasonal version 1 (ACCESS-S1) 12 times up 217 days. Leave-one-year-out testing results illustrate that has normally higher accuracy skill, measured mean absolute error CRPS respectively, than QM. substantially improves ACCESS-S1 raw does always outperform climatology, benchmark SCFs. Many more research efforts required modelling skilful