[edit]
Weather4cast at NeurIPS 2023 – Data Fusion for Quantitative Hi-Res Rain Movie Prediction under Spatio-temporal Shifts
Proceedings of the NeurIPS 2023-2025 Competition Track, PMLR 276:90-105, 2026.
Abstract
Now in its third year, Weather4cast 2023 again advanced modern algorithms in AI and machine learning through a highly topical interdisciplinary challenge at the NeurIPS Competition Track. Predicting hi-res rain radar movies from multi-band satellite sensors required data fusion of complementary signal sources and video frame prediction with super-resolution techniques. This year, quantitative prediction of rain rates was assessed using a balanced multi-class metric. Accurate rain predictions are ever more critical for everyone, with climate change increasing the frequency of extreme precipitation events. Weather4cast models thus have a particular impact for the many regions on Earth where costly weather radar data are not available. We here present highlights and insights emerging from the competition. To reward models extracting relevant mechanistic patterns reflecting the underlying complex weather systems, our evaluation tested their responses to spatio-temporal shifts. Algorithms needed to forecast 4 h or 8 h of ground-based hi-res precipitation radar from lo-res satellite spectral images in a unique cross-sensor prediction challenge, and were evaluated within and across a variety of regions on Earth with diverse climate and different distributions of heavy precipitation events. Robustness to distributional shifts over time was achieved by testing predictions on data one year after the training period. Notably, the capability of models to correctly capture spatio-temporal patterns qualitatively varied considerably and was not always in line with leaderboard scores. This suggests that more sophisticated metrics will be required to accurately reflect the effectiveness of AI models in corresponding meteorological applications. Weather4cast will continue to explore the expansive data set and advance benchmark challenges in the field en route to validated foundation models for weather.