WeatherForecast

English · 简体中文

Hourly ERA5-derived weather fields for lower-cost forecasting experiments. Both subsets store unnormalized float32 values and continuous sequences, not pre-extracted windows.

Datasets

	TyphoonTracking	GlobalForecast
Related experiment	Met2Net Table 5	Met2Net Table 4
Time between consecutive frames	1 hour (source setting)	1 hour (UTC timestamps included)
Grid	32 × 32, approximately 1°	32 × 64, 5.625° spacing
Coverage	110–142°E, 12.25–44°N¹	Global²
Array shape	(4, T, 32, 32) — C,T,H,W	(T, 5, 32, 64) — T,C,H,W
Channel order and units	msl (Pa), u10 (m/s), v10 (m/s), t2m (K)	u10 (m/s), v10 (m/s), t2m (K), tcc (fraction), r at 1000 hPa (%)
Data files	train_32_32.npz, val_32_32.npz, test_32_32.npz	train.npy, val.npy, test.npy
Statistics / coordinates	NPZ keys: data, stats; no coordinates	train_stats.json, coordinates.npz

Split	TyphoonTracking: years¹ / frames	GlobalForecast: years / frames
Train	2017–2021 / 43,824	2010–2015 / 52,583
Validation	2022 / 8,760	2016 / 8,783
Test	2023 / 8,760	2017–2018 / 17,519

¹ Regional bounds and years follow Met2Net's reported setup; source NPZ files contain no coordinates or timestamps for independent verification. Latitude direction is not established.

² Global cell centers: latitude −87.890625° to 86.484375°, south to north; longitude 1.40625° to 355.78125°E. coordinates.npz provides lat, lon, and time_train/val/test. Each split ends at 22:00 UTC on its final day because earlier preparation removed its last hour.

Preparation and use

• TyphoonTracking: 128 × 128 → 32 × 32 by non-overlapping 4 × 4 arithmetic means. Normalize with (raw - min) / (max - min + 1e-8) using the NPZ stats dictionary (min/max/mean/var per channel; reading it requires allow_pickle=True).
• GlobalForecast: WeatherBench's 1.40625° grid was first sampled every second point to 64 × 128, then reduced by 2 × 2 arithmetic means. Normalize with (raw - mean) / std using per-channel [mean, std] in train_stats.json (ddof=0). This is not the official WeatherBench 5.625° product.
• Both use float64 averaging followed by float32 storage, without area weighting or temporal downsampling. Statistics are recomputed from the downsampled training set only and reused for validation/test. Do not refit or automatically clip validation/test values.
• A 12-input / 12-target window covers hours −11…0 → +1…+12. Build windows within each split. Use NumPy to read the files; GlobalForecast NPY files support mmap_mode="r".

Limitations: these are lower-resolution derivatives, not the original Met2Net benchmarks. TyphoonTracking includes non-typhoon periods and has no track labels. Coarse fields smooth cyclone structure; field error does not measure track accuracy.

Sources and terms

ERA5 single levels · ERA5 pressure levels · WeatherBench (Rasp et al., 2020) · Met2Net (Li et al., ICCV 2025) · Authors' ERA5-Cropped release

Independently prepared on 2026-10-23. Contains modified Copernicus Climate Change Service information. Upstream data terms apply; this repository does not relicense the data.