Leveraging satellite imagery and yield records to assess in-season maize field functionality with machine learning classification

• ML-based classification of maize fields in-season functionality used Sentinel-2 imagery and yield data. • Integration of spatial, temporal, and spatio-temporal features enhanced classification accuracy. • OvR binary classification outperformed multi-class approaches, validated via LOYO cross-validation. • Feature selection, scaling, and ML methods had minimal impact on the overall model performance. • Spatio-temporal and SWIR features improved early-stage classification, apparently capturing soil moisture variability. This study introduces a framework for assessing in-season maize field functionality from 5 to 70 days after sowing (DAS) using machine learning (ML) classification of Sentinel-2 satellite imagery. Yield records from 123 maize fields (2018–2021) served as proxies for functionality levels. Unlike conventional methods relying solely on spectral vegetation indices, the presented framework integrates spatial, temporal, and spatio-temporal feature types. The methodology employed multi-class and multiple one-vs-rest (OvR) binary classifications using various machine learning algorithms, validated through a leave-one-year-out (LOYO) cross-validation strategy to ensure robust real-world applicability. Combining all 4 feature types, consistently improved classification performance across 14 consecutive pentads after sowing by an average of 5.3 % and a maximum of 10 % (F1-score) compared to spectral vegetation indices alone. The model achieved reliable functionality classification even during early growth stages, where spatial features and shortwave infrared indices played crucial roles, apparently revealing initial soil moisture variations. Multiple OvR binary classifications outperformed multi-class classification, while no significant differences emerged between data preparation methods or ML algorithms. While all feature types and spectral ranges contributed to the classification, features’ importance levels shifted throughout growth stages, reflecting changing spectral and feature-type contributions, which are further discussed. The classification framework has been integrated into a web-based decision support tool, providing farmers with real-time functionality monitoring. To better understand the contribution of spatial and temporal features to field functionality classification, future research should incorporate additional yield records and other relevant parameters. By using the tool, farmers will be able to contribute new yield data, facilitating continuous learning and improvement of the model.