Probabilistic prediction of climate extreme events with deep learning

Climate extreme events pose significant societal and environmental challenges, yet their prediction remains complex due to their rare occurrence and inherent variability. We present a novel probabilistic deep learning framework that predicts multiple climate extremes directly from simple variables, namely monthly temperature and precipitation. Our approach employs a conditional generative adversarial network with a modified U-Net architecture, incorporating self-attention mechanisms and fully-residual blocks to capture long-range spatial dependencies and provide precise estimations. The model is trained using a combination of adversarial, perceptual, physical, and frequency losses, along with an extensive data augmentation pipeline designed explicitly for gridded climate data. We show that our approach achieves superior performance compared to different baselines in predicting nine different climate extreme indices, including droughts, temperature extremes, heat waves, cold waves, and precipitation and snow extremes. Importantly, our framework provides uncertainty estimates, essential for decision-making in climate adaptation strategies. Through comprehensive ablation studies, we show the relative importance of different architectural components and training strategies. Our results suggest that deep learning can effectively bridge the gap between monthly climate variables and extreme event prediction, offering a computationally efficient alternative to traditional climate modeling approaches while maintaining physical consistency and providing uncertainty quantification.