Improved Seasonal Precipitation Forecasts for the Blue Nile Basin: A Deep Learning Approach

Abstract

Seasonal precipitation forecasts are essential for climate-sensitive sectors such as agriculture and water management in East Africa. However, the application of seasonal forecasts at regional scales requires post-processing due to systematic errors and insufficient spatial resolution to capture local characteristics. Yet current statistical methods have remaining limitations in terms of spatial consistency and the representation of extreme events. Here, we propose a deep learning approach, Seasonal AFNOCast, based on an adaptive Fourier Neural Operator architecture, to bias-correct and downscale SEAS5 precipitation forecasts for the Blue Nile Basin, a transboundary catchment in Ethiopia and Sudan. We evaluate Seasonal AFNOCast alongside the established statistical method, Bias Correction and Spatial Disaggregation (BCSD), using forecasts from 2017–2023. Results show that both methods substantially improve precipitation distributions, spatial patterns, and the Continuous Ranked Probability Skill Score (CRPSS) of approx. 0.3 compared to raw SEAS5. Despite only modest improvements over climatology across the entire evaluation period (CRPSS approx. 0.03), both methods show clear skill enhancements during the months March to May (MAM), a highly variable yet operationally critical season for decision-making. While onset predictability remains challenging at a seasonal scale, even after post-processing, this study identifies key differences in the application of the post-processing methods: BCSD performs best at short lead times, whereas Seasonal AFNOCast maintains higher skill at longer leads and indicates an improved representation of high-intensity rainfall and spatial frequency characteristics. Moreover, Seasonal AFNOCast generates forecasts 5–20 Rebecca Wiegels et al. Deep Learning Improved Seasonal Forecasts times faster than BCSD, making it particularly suitable for operational contexts. Our findings show that deep learning can complement and extend conventional post-processing, improving seasonal forecasts for subsequent applications and supporting hydrological and agricultural decision-making where representation of extreme events and spatial consistency, as well as computational efficiency, are critical.