Machine learning technology for early prediction of grain yield at the field scale: A systematic review

•Assessed the adoption of machine learning to predict field-scale grain yield.•Richness of data collection, preprocessing, training and evaluation revealed.•Prediction horizon is fundamental to the interpretation of results from studies.•Lack of consensus regarding performance metrics undermines integration of evidence.•Enhanced reporting suggested for promoting the accumulation of knowledge. Machine learning (ML) has become an important technology for the development of prediction models for crop yield. Predictive modeling using ML is rapidly growing as research addresses early predictions versus predictions shortly before harvest, predictions at the scale of field or region, and predictions for different types of crops. This great diversity of prediction tasks requires a proper choice of specific ML techniques to attain high levels of performance. Therefore, this review focuses on a distinct prediction task and aims to provide task-specific insights into the adoption of ML. The objective of our research is to investigate ML approaches for the early prediction of grain yield at the field scale. We identified studies published between 2014 and 2021 through a systematic search in Scopus and Web of Science for journal articles and a retrieval of analogous articles from three previous reviews. The study selection process included screening, full-text assessment, and data extraction by two independent coders. Of 924 unique records identified in the search and retrieval, 157 full texts were assessed for eligibility, and 46 studies met all inclusion criteria. The results paint a comprehensive picture of the ML techniques used, revealing the richness of data collection, preprocessing, model training, and model evaluation. Specifically, the results highlight (1) a wide range of prediction horizons from a few weeks up to more than eight months before harvest; (2) a large set of input data representing weather, crop management, site characteristics, and vegetation properties; (3) a low level of adoption of feature selection methods to enhance performance; (4) some lack of information on the size of the training and test sets required to assess their suitability; and (5) heterogeneity in the reporting of performance metrics that hinders the comparison and integration of evidence from individual studies. To overcome barriers to the accumulation of evidence, we suggest recommendations for enhanced reporting and building greater consensus regarding the most appropriate