Enhancing WorldCereal crop calendars with land surface phenology and machine learning

Accurate crop calendars are critical for understanding global agricultural phenology, mapping crop types, improving yield forecasts, and addressing food security challenges in the context of climate variability. This study presents significant advancements in crop calendar modeling, building upon the WorldCereal system by integrating Land Surface Phenology (LSP) metrics derived from MODIS AQUA NDVI time series with ERA5-Land climate parameters and machine learning algorithms, particularly XGBoost. Our new approach introduces key enhancements, including densification of training data in underrepresented regions, explicit modeling of dormancy periods for winter cereals, and the synergistic use of climate and phenological predictors to refine Start of Season (SOS) and End of Season (EOS) estimates across diverse agro-ecological zones for eleven crops types explicitly excluding rice, which falls outside the scope of the WorldCereal project. The model was validated against harmonized ground-truth data, remote sensing-based phenology retrievals, and expert assessments from agricultural monitoring institutions, achieving R 2 values of 0.94 (SOS) and 0.92 (EOS) for summer crops, with an RMSE of 18 and 22 days, respectively. For winter crops, the model reached R 2 values of 0.90 (SOS) and 0.92 (EOS), with RMSE values of 29 and 28 days. The inclusion of dormancy modeling significantly improved the representation of winter cereal cycles, reducing previous overestimations in high-latitude regions such as Canada, Northern Europe, and Central Asia. Additionally, enhanced spatial resolution addressed gaps in South America, Africa, and Asia, refining SOS and EOS timing in monsoon-driven cropping systems. Expert validation from institutions such as INPE (Brazil), NASA Harvest, and the Chinese Academy of Sciences further refined estimates, correcting anomalies in China and southern Brazil, ensuring alignment with observed agricultural cycles. The findings demonstrate the added value of combining remote sensing-derived phenology and climate data with XGBoost-based machine learning to improve crop calendar precision, optimize planting and harvesting schedules, and mitigate risks from climate variability. This study establishes a robust, data-driven tool for agricultural monitoring, supporting yield forecasting and global food security initiatives. • Enhanced WorldCereal global crop calendars using MODIS NDVI and ERA5-Land. • Modeled winter crop dormancy to improve phenological accuracy at scale. • Used XGBoost to fuse climate and phenology features for SOS/EOS prediction. • Validated against expert assessments and external geo-referenced datasets. • Achieved R 2 > 0.90 and RMSE <30 days across crop types and regions.