A scalable machine learning approach for predicting wheat growth stages with a large national dataset

Context Accurate wheat growth stage predictions are important for efficient crop management practices, such as when to apply chemical inputs or fertilise. Recent studies have developed accurate machine learning (ML) approaches for predicting the growth stages of wheat and other crops, but these have generally focused on a few key stages and used limited datasets, restricting comprehensive validation across diverse growing seasons and/or regions. This makes it difficult to test their scalability, which is a critical consideration for real-world application. The National Variety Trials (NVT) program represents a key opportunity, providing observations of Zadoks stages since 2005 across the Australian grain belt. Aim To develop a scalable, data-driven approach for predicting wheat growth stages using a national dataset and ML. Methods The dataset contained over 80,000 wheat Zadoks stage observations from the NVT program from 2005 to 2023 across 169 sites in Australia. Models were developed with XGBoost, using 11 weather, remote sensing (RS), genetic and crop management features. Three experiments were designed to evaluate models: 70:30 split, leave-one-year-out (LOYO) and leave-one-site-out (LOSO). The optimal spatial extent was determined by comparing national, regional and subregional models, and a null model was developed to assess quality of predictions if only using features related to temperature. Results All three spatial extents tested yielded strong results, but the national performed best overall. It showed high accuracy across all three experiments, with strong agreement between observed and predicted Zadoks stages (00−99) (Lin’s concordance correlation coefficient LCCC = 0.76–0.80), minimal error (RMSE = 5.9–6.8 stages), and 58–66 % accuracy ±5 Zadoks stages across the three experiments. This model also consistently outperformed the null model, demonstrating that including non-temperature-related features (e.g. solar radiation, variety) led to more accurate growth stage predictions. Conclusions No other known studies have used such a comprehensive dataset of growth stage observations, both in size and spatiotemporal coverage, to model crop growth stages. The use of this dataset enabled the development and validation of a model that is both accurate and scales well to unseen years and sites. This study therefore highlights the potential for a ML-based operational tool to support crop monitoring across diverse growing seasons and regions. Future work could explore incorporating more RS features and more observations of underrepresented Zadoks stages (e.g. seedling growth).