CycloneNet: A Reproducible Pipeline and Leakage-Safe Two-Basin Dataset for Tropical-Cyclone Rapid-Intensification Analysis

This record contains the CycloneNet v3 preprint and the v2 dataset package. The repository code snapshot (tag v3.0.1) is not attached here; it is available at https://github.com/estefano-ferreira/cyclone-net/archive/refs/tags/v3.0.1.zip (also under Related works). SUPERSEDES v1.0.0, v1.0.1 AND v2.0.0Section 9 of the enclosed preprint documents every correction, with the reproducible numbers that replace the withdrawn ones. In short: v1.0.0 / v1.0.1 — ROC-AUC 0.97, recall 0.92, "sub-pixel spatial accuracy", "~26 km mean spatial error", "18 hurricanes (1989–2024)": withdrawn for scope, not provenance. Do not cite.These numbers are reproducible — they come from validation_report.txt, released in both v1 records, and recomputation from the accompanying cyclonenet_scientific.csv reproduces every one exactly. What does not hold is the generalization: the benchmark comprises 58 track points from 19 named Atlantic hurricanes, 26 of them RI-positive — a 44.8% prevalence, reported without confidence intervals or a declared train/test separation. It is not comparable to any measurement on the full archive (16,780 events, 4.8% prevalence). v2.0.0 — ROC-AUC 0.83, recall 0.905 on 2,193 test samples: not reproducible from the public artifacts. Do not cite.The v2 release covered only 2020–2023 (972 events; test split 155 samples); the reported metrics were computed on data that release did not contain. The reproducible replacements, from the full 1980–2023 archive: ROC-AUC 0.796 [0.753–0.837], PR-AUC 0.251 [0.179–0.331], recall 0.852 on a frozen test split of 2,679 events — metrics of a model now retired, standing as historical record only. Also withdrawn: the FuelMap spatial-attribution claim was never externally validated and remains unsupported — co-location tests against audited TCHP peaks did not distinguish it from a storm-centre baseline. The "physics-guided" framing is retired: the released v2 code left every training.physics.* weight at 0.0, so the model trained as a plain CNN. The 3D-CNN architecture is retired by a pre-registered verdict. The validated contribution is the dataset and the auditable pipeline. CONTENT16,780 valid events from 992 tropical cyclones (578 East Pacific / 414 North Atlantic by genesis basin), 1980–2023, June–November, over [60°N, 140°W, 0°N, 20°W]. Each event is a 6-hourly IBTrACS track point with its 24-hour preceding ERA5 history, stored as a float32 cube of shape (40, 40, 5, 14): 40×40 grid window (~10°×10°) centred on the storm, five 6-hourly time levels (t0, −6h, −12h, −18h, −24h), 14 channels. LABELS (v2, strict-temporal, tri-state)799 RI positives (ΔV24 ≥ 30 kt, Kaplan & DeMaria 2003), 15,962 negatives, 19 NULL. NULL is a label state meaning UNDEFINED — the exact 24 h temporal partner does not exist (storm dissipation or reporting gap). It must not be coerced to 0, and NULL events must be excluded from RI-classification tasks. SPLITSTrain/validation/test assigned at the storm level (SID) by SHA-256 hash-determinism with a frozen override map; no storm crosses splits. Normalization statistics computed on the training split only. The frozen test split (2,679 events; 112 RI+, 6 NULL) is distributed for reproducibility. The design controls train/test non-independence through storm identity; it does not control the temporal axis (storms from all decades appear in all splits). Per-event SID and timestamp are included, so year-based or leave-one-year-out splits can be constructed directly. PROVENANCEPer-window SHA-256 manifests, label-correction audit (label_diff_v1_v2.csv, 32,989 rows), basin-repair record. The labeling chain is byte-reproducible from raw IBTrACS (32,989/32,989 rows) via the replication-gate script in the code snapshot. CRITICAL READER NOTEThe string "NA" in the basin field means North Atlantic — it is a categorical value, not a missing value. When reading any distributed CSV with pandas, use keep_default_na=False, na_values=[""]. Without this, "NA" is silently coerced to NaN and basin assignments are corrupted. This bug class occurred six times during this project's own development. FILESThe preprint (PDF) and the data package — 46 files, 6.03 GiB: 44 per-year cubic shards, 1 metadata archive, 1 checksums file (see DATA_DICTIONARY.md and TECHNICAL_VALIDATION.md inside the metadata archive). The code snapshot is not attached; it is at the tag URL above. LICENSEDataset: CC BY 4.0. Code: MIT (see LICENSE in the repository snapshot linked above). Redistribution of the dataset must preserve the third-party attributions below (see the NOTICE file). ATTRIBUTIONS (mandatory) ERA5: Contains modified Copernicus Climate Change Service information (1980–2023). Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains. Hersbach, H., Bell, B., Berrisford, P., et al. (2018), ERA5 hourly data on single levels (doi:10.24381/cds.adbb2d47) and on pressure levels (doi:10.24381/cds.bd0915c6), Copernicus Climate Data Store. IBTrACS: International Best Track Archive for Climate Stewardship (IBTrACS), Project Version v04r00, NOAA National Centers for Environmental Information, doi:10.25921/82ty-9e16 (Knapp, K. R., et al., 2010, Bull. Amer. Meteor. Soc., 91, 363–376). IBTrACS follows the World Data Center for Meteorology policy of full and open access; WMO Resolution 40 guides commercial use. NOAA/NCEI provide the data without warranty. ADT (auxiliary, 761 events, 2020–2023): Generated using E.U. Copernicus Marine Service information. CODEhttps://github.com/estefano-ferreira/cyclone-net (MIT) Contact: [email protected]