Considering that studies handling machine learning approaches for data analysis of OFEs are increasing recently, our study may be helpful for enhancing the model inter- pretability of crop yield response to variable rate application of fertilizer and seed. We describe the risk of overfitting and enhanced model explainability by introducing spa- tial surrogate covariates in a machine learning model. For a better prescription map through on-farm experimentation, machine learning, such as random forest (RF), has been proposed to model the site-specific yield response function. Geographical coordinates are sometimes used as covariates to mimic soil variations. However, tree-based models tend to overfit, which potentially hinders the understanding of causal relationships. The aim of this study was to explore the pos- sibility that the explainability of machine learning can be enhanced by using Moran’s eigenvector maps (MEMs) as surrogate covariates. Site-specific wheat yield responses to fertilizer were modeled using RF with either coordinates or 1–10 MEMs as covariates, followed by a sensitivity analysis. Several sites exhibited negative and unreliable yield responses to fertilizer when the coordinates were used as covariates. At the same time, in the model with MEMs, yield was almost always positively responsive to fertilizer, which was a reliable outcome. Automation MEM selection should be examined in fur- ther studies.