Prepared inputs, model checkpoints and case-level results for cherry flowering prediction at 23 Korean stations

This dataset supports reproducible evaluation of cherry first-flowering prediction at 23 stations in South Korea. It contains prepared inputs, station metadata, case-level observations and predictions, corrected learning-model features and checkpoints, evaluation settings, figures, and reproducibility code. The primary comparison includes a historical station-median baseline, three process models, two conventional machine-learning methods, and three neural networks. Development evaluation follows time-ordered prediction for 2010–2022. For the primary 2023–2026 re-evaluation, model fitting and hyperparameter selection used only data through 2022. Results for 2023–2026 had already been examined before the original photoperiod implementation was corrected and the baseline eligibility rules were revised. The original photoperiod implementation used the within-season sequence index in place of calendar day-of-year. The calendar-correction rules were fixed before generating the corrected predictions, and later-period scores were not used for hyperparameter selection or to choose the correction rule. The 2023–2026 results therefore constitute a retrospective temporal re-evaluation, not a previously untouched independent test of a prespecified analysis procedure. Primary comparisons use common sets of 265 and 83 station–year cases, respectively, alongside model-specific available cases. A fixed-parameter retrospective sensitivity used fully observed post-issue weather, assuming complete knowledge of those weather inputs. On the same 83 cases in 2023–2026, mean absolute errors fell from 4.14 to 2.17 days for the parallel model (PA) and from 4.46 to 2.64 days for the thermal time model (TT). Both reductions exceeded the 0.99-day spread among the nine primary methods under simulated issue conditions in that period. This comparison describes sensitivity to weather information; it is neither a guaranteed upper bound on improvement nor a measurement of operational forecast skill. Full-bloom results are provided separately. Full raw hourly weather archives and some process-model refitting inputs are excluded; reproduction limits are stated in the README. Phenological observations, meteorological observations, and station information were obtained from the Korea Meteorological Administration (KMA). The observations were processed for this study. Official source catalogues are:https://www.data.go.kr/data/15139432/openapi.dohttps://www.data.go.kr/data/15139437/openapi.dohttps://www.data.go.kr/data/15139439/openapi.do Licenses apply to separate components. Original research contributions, documentation, figures, and numerical model states are licensed under CC BY 4.0; original source code is licensed under MIT. KMA source materials and their source components in processed and mixed files retain Korea Open Government License Type 1 attribution conditions. Natural Earth coastline data remain public domain. See RIGHTS.md and DATA_SOURCES.md for scope, attribution, and third-party notices. This work was supported by the Rural Development Administration, Republic of Korea, under project PJ01778803.