Subseasonal-to-seasonal forecasts through self-supervised learning

Sub-seasonal to seasonal (S2S) weather forecasts play a crucial role in guiding decision-making processes related to agricultural planning, energy management, and disaster mitigation. Operated on time scales spanning weeks to months, these forecasts distinguish themselves from short-term predictions in two key aspects: (i) the atmospheric dynamics on these timescales are accurately described only through statistical means, and (ii) these dynamics exhibit large-scale phenomena in both spatial and temporal dimensions. Despite the success of deep learning (DL) in short-term weather forecasting, DL-based S2S predictions face challenges arising from limited training data and significant predictability fluctuations due to varying atmospheric conditions. To enhance the reliability of S2S forecasts by incorporating the latest DL advancements, our proposal involves the application of the masked auto-encoder (MAE) framework. This framework aims to learn comprehensive representations of large-scale atmospheric phenomena from high-resolution global data. Beyond assessing the suitability of these learned representations for S2S forecasting, our investigation extends to their potential to account for climatic phenomena, such as the Madden-Julian Oscillation, recognized for enhancing predictability on S2S timescales.