Seasonal AFNOCast: A Deep Learning Approach for Enhanced Regional Seasonal Predictions

Regionalized seasonal forecasts allow improved decision making, particularly when applying the meteorological forecasts to sectors such as agriculture or water management. In regions like the Blue Nile Basin, a transboundary catchment in East Africa, reliable seasonal predictions are crucial for addressing local needs due to the complex topography combined with high dependency on water resources.

In this study, we introduce Seasonal AFNOCast, a Deep Learning (DL) approach designed to bias-correct and downscale global seasonal forecasts (SEAS5). The objective is to provide a computational efficient approach that provides reliable ensembles, realistic and skillful predictions at a daily and monthly scale. The regionalized forecast provides 51 ensemble members with a 215 day forecast horizon of a spatial resolution of approximately 9 km.

Seasonal AFNOCast is a DL network that applies a specific type of transformer, called the Adapted Fourier Neural Operator (AFNO), in combination with an ensemble-member-specific architecture. The network is trained with the ERA5-Land reanalysis product as reference using an ensemble specific loss function. Its performance is evaluated against Bias-Correction and Spatial Disaggregation (BCSD), a well-established statistical baseline method for post-processing global seasonal forecasts. The evaluation includes comprehensive skill metrics such as the continuous ranked probability score (CRPS), normalized rank histograms, and precipitation-specific metrics, along with qualitative analyses.

Our analysis demonstrates that Seasonal AFNOCast delivers skillful regionalized seasonal predictions that are comparable to, and in specific cases outperform, state-of-the-art statistical methods. These findings underscore the potential of DL-based post-processing of seasonal forecasts, particularly in challenging regions like the Blue Nile Basin.