Applied AI research
EastAsiaClimateExtremes Turns East Asian Heatwaves, Heavy Rainfall, and Marine Heatwaves into Weekly AI Labels
The APEC Climate Center has released an East Asian extreme climate dataset aligned with ECMWF subseasonal reforecasts, covering gridded labels, event intensity, and analysis code. It reduces the preprocessing cost of creating AI prediction targets, but its 1.5-degree grid, biases in reanalysis data, and unclear licensing status limit direct deployment.
Submitted on September 8, EastAsiaClimateExtremes addresses a common data gap in AI research on extreme weather in East Asia: although ERA5, OISST, and numerical forecasts provide continuous meteorological fields, researchers still have to convert temperature, precipitation, and sea surface temperature data into event labels suitable for training and validation. The new dataset places anomalous heat, heavy rainfall, and marine heatwaves on a common 1.5-degree grid spanning 21°N to 48°N and 114°E to 141°E. It provides daily and weekly time series, climatological means, and 90th- and 95th-percentile thresholds.
The processed data indicate more than whether a threshold was exceeded: they also preserve the number of extreme days, peak intensity, and cumulative impact. Events are defined using a minimum duration and an allowed gap. For example, `D3G5` denotes an event lasting at least three days with gaps of up to five days permitted during the event. Records include start and end dates, duration, and mean and peak intensity. ERA5 coverage spans 1940–2024, while OISST covers 1982–2024. The authors also place the labels and the 2016 and 2024 versions of ECMWF subseasonal reforecasts in the same spatiotemporal framework. An example workflow compares results for lead week 3—that is, days 15–21 after forecast initialization.
The dataset’s technical value lies in separating “forecasting mean meteorological fields” from “identifying low-frequency, high-impact events,” turning them into reproducible supervised-learning and evaluation tasks. The GitHub repository provides NetCDF and NumPy outputs, along with notebooks for time-series analysis, heatmaps, long-term trends, and two-dimensional statistics. These resources are suitable for building classification, probabilistic forecasting, or hybrid dynamical–AI baselines.
However, this is not a new forecasting model, nor does it demonstrate that AI can outperform ECMWF. The labels are derived primarily from reanalysis data and percentile-based rules, so they inherit biases from both the source data and the event definitions. The 1.5-degree resolution is also insufficient to represent many localized heavy-rainfall events. The repository currently does not clearly specify software and data licenses, so teams should seek confirmation before redistribution or commercial use.