Deep Learning-Based Extreme Heatwave Forecast

Because of the impact extreme heat waves and domes on society biodiversity, their study is a key challenge. We specifically long-lasting waves, which are among most important for climate impacts. Physics driven weather forecast systems or models can be used to occurrence predict probability. The present work explores use deep learning architectures, trained using outputs model, as an alternative strategy heatwave. This new approach will useful several scientific goals include model statistics, building quantitative proxy resampling rare events in models, change, should eventually forecasting. Fulfilling these implies addressing issues such class-size imbalance that intrinsically associated with event prediction, assessing potential benefits transfer address nested nature (naturally included less ones). train Convolutional Neural Network, 1,000 years outputs, large-class undersampling learning. From observed snapshots surface temperature 500 hPa geopotential height fields, network achieves significant performance forecasting heatwaves. able them at three different levels intensity, early 15 days ahead start (30 end event).