Abstract: To improve the forecasting accuracy of early rice heading date across diverse meteorological regimes, we developed a scenario-based model that integrates effective accumulated temperature (EAT) with support vector machine (SVM), trained on 30 years (1994-2023) of agrometeorological records from Jiangxi Province, China. The applicability and predictive performance of this model were then evaluated. Results showed that the EAT model, calibrated with 15 ℃ base temperature and adjusted for latitude, achieved high overall accuracy for early rice heading date prediction (mean absolute error (MAE)=3.1 d, accuracy=91.6%), and was particularly suitable for long leadtime forecasting under normal weather from tillering to heading. However, when this period experienced any of four abnormal weather regimes—sustained low/warm temperature, stage strong cooling/warming events, the EAT error significantly increased. The SVM model that sequentially ingested key meteorological factors, including the number of cloudy days in mid-May, temperature factors, and sunshine hours from mid-May to early June, as well as sunshine hours and cloudy (rainy) days in mid-June, effectively reduced these errors: among samples where the EAT errors were ≥3 d, SVM lowered MAE from 5.2 d to 3.3 d (a 36.5% reduction). The correction was most pronounced for prolonged low-temperature episodes, with the average forecast error being reduced from 5.7 d to 2.4 d (a 57.9% reduction). Accordingly, we proposed a scenario-adaptive strategy “EAT for normal conditions and SVM correction for abnormal conditions”, that maintains EAT’s advantage under normal weather while improving accuracy under climate extremes, providing an effective solution for precise forecasting of crop development stage under complex climatic conditions.

Key words: early rice, heading date forecast, support vector machine, effective accumulated temperature