Weather-based prediction of black locust (Robinia pseudoacacia L.) honey yield and quality: a reproducible computational system with operational deployment

Abstract

Black locust (Robinia pseudoacacia L.; marketed as “acacia”) honey production in South Korea—accounting for over 70% of national honey output—has been increasingly affected by weather-driven disturbances, resulting in substantial production declines, notably in 2018 and 2020 ( Kim et al., 2022 ). Such episodes have been broadly attributed to impacts of climate change; however, a diagnostic framework identifying which weather variables, and at which phenological stages, drive the simultaneous decline in yield and quality has remained unavailable. Existing approaches rely either on uninterpretable machine-learning models that require large training data unavailable at the farm scale, or on single-variable rules-of-thumb that fail under multivariate weather conditions. Therefore, we present a reproducible computational system that addresses this gap through three integrated components: (i) two interpretable hierarchical weather indices—the Honey Harvest Weather Index (HHWI) for yield and the Honey Moisture Weather Index (HMWI) for moisture content—derived from five meteorological variables (maximum temperature, precipitation, relative humidity, wind speed, and solar radiation) across three phenological stages (bud formation, pre-flowering, and flowering); (ii) a generalized raster pipeline producing spatially-explicit predictions from gridded meteorological data; (iii) an operational forecast module generating real-time decision-support maps via inverse-distance interpolation. The system was validated using 570 farm-year records from 95 black locust apiaries (2020–2025) in South Korea and benchmarked against seven machine-learning alternatives. HHWI achieved a cross-validated R2 value of 0.395 (RMSE = 16.85 kg/colony) using three effective parameters and outperformed XGBoost in terms of cross-validated stability, exhibiting a 15-fold reduction in the overfitting gap. HMWI achieved a cross-validated R2 value of 0.668, with a moisture-grade classification accuracy of 96.0% (within ± 1 grade, Cohen’s κ = 0.41). This demonstrates its operational feasibility for weather-based quality prediction. Retrospective decomposition of the 2020 production minimum (18.4 kg/colony) indicated that approximately 60% of the deficit relative to the 2024 peak year (60.5 kg/colony) was attributable to maximum temperature during the pre-flowering stage. Cumulative temperature contributions across all phenological stages accounted for 80% of the deficit, transforming broad climate-change attribution into actionable diagnoses of specific weather variables and phenological stages. The data-driven breakpoints converged with published physiological thresholds for Apis mellifera , supporting the mechanistic interpretation. The complete system—including analysis script, raster pipeline, forecast module, and 22-checkpoint verification—was released as an open-source code, providing beekeepers and policymakers with a deployment-ready, weather-based decision-support system that requires only routine meteorological data. All codes and anonymized data are available on Zenodo ( https://doi.org/10.5281/zenodo.20612458) .