Neglecting spatial autocorrelation leads to underestimation of the error in the development of sugarcane yield models

Sugarcane yield models, as most crop predicting models, are subject to the existence of spatial autocorrelation between observations. In this work, we used machine learning techniques to generate sugarcane yield models by splitting blocks of data, grouped by distance, in training and test sets in the cross validation phase, in contrast with separating single observations, as if they were independent. Although models generated using blocks of data led to a better estimation of the error in new data, both approaches generated similar error values.