What makes on-farm experimental data suitable for data-driven decision-making? Implications of trial design and spatial distribution of field data for machine learning models

Abstract Purpose On-farm experimentation (OFE) plays a key role in underpinning data-driven agronomic decision-making. For example, machine learning (ML) models can predict optimal nitrogen (N) fertiliser rates using trial crop responses alongside soil, plant, and climatic data. However, it is unclear how different OFE strategies, trial designs, and their consequences for the spatial distribution of field datasets, impact the development of such models. This work sought to investigate how trial design influences spatial autocorrelation in OFE data and the impact of this on model training. It was also of interest to explore whether tailoring OFE programs and ML models to specific regions might improve their performance compared to those generated for large geographic areas. Methods Using 21 N strip trials across Australia, ML models were developed to predict optimal N rates under different scenarios of data autocorrelation and geographic coverage. Results Spatial autocorrelation in OFE data had negligible impact on model performance. At the same time, models trained with fewer non-correlated observations showed similar performance to models trained with thousands of autocorrelated observations. This suggests that less replicated field trials providing more independent observations might be preferable – for their simplicity and pragmatism – to highly replicated, whole-field trials which generate highly autocorrelated field data. The results also indicate that regional models may perform better than global models. Conclusion Overall, to improve both the quantity and quality of OFE data for ML models used to underpin a mid-season N fertiliser decision, prioritising a greater number of simpler experiments (less replicated strip or plot trials) across a region is likely to be more effective than increasing field coverage of individual experiments using highly replicated whole-field designs focused on field-specific models.