A Methodological Critique of Machine Learning in Tanzanian Agriculture: Rigor, Reproducibility, and a Roadmap for Real-World Impact

Machine learning (ML) holds immense promise for transforming Tanzanian agriculture. However, a comprehensive review of over 20 peer-reviewed studies reveals prevalent methodological shortcomings. Despite often reporting high accuracy, many models are hindered by: (1) training on data that does not accurately represent real-world conditions, (2) insufficient validation methods (such as the absence of spatial or time-series cross-validation), (3) the use of inappropriate metrics (e.g., relying on accuracy for problems with imbalanced datasets), and (4) a notable decline in performance when moving from experimental settings to practical agricultural applications. These critical issues undermine the models' validity and severely restrict their tangible impact. This paper outlines a strategic roadmap for researchers, academic journals, and funding bodies to enhance the rigor, reproducibility, and real-world applicability of agricultural ML research within Tanzania.

Working paper