Applying machine learning digital soil mapping techniques on farm scale for use in precision agriculture

Abstract

This study aimed to evaluate the industry-standard method of ordinary kriging (OK) used for creating soil chemical property maps and then also to investigate the potential of using machine learning-based digital soil mapping techniques to create the same maps for use in precision agriculture (PA). To do this, two fields in the Gerdau area in the North West of South Africa were selected, and soil chemical analyses for them on a 2-ha grid were obtained from NWK. Using similar processes to the industry standard, OK maps for pH, P, K, Ca, Mg and Na were produced, and the accuracy thereof was tested using leave-one-out cross-validation. The semivariograms of each property were also investigated to determine the optimal sampling density. After this, the same sets of maps were created using the ML DSM techniques Cubist and Random Forest (RF), the accuracy of which was tested by splitting the data into a training and validation set. The ML DSM maps were then also compared to the OK maps through visual inspection. It was found that soil properties varied between Fields and each other. In Field 1, only the OK map of Mg was somewhat accurate enough for use in PA. More success was found in Field 2 with the OK maps of pH, K and Mg being accurate enough for use in PA. Based on the semivariograms of these properties, a smaller inter-sample range is needed than the 2-ha grid provides. Based on the results of this study, it is recommended that sample grids no larger than 70 by 70 meters be used, in line with the findings of Brouder and Morgan (2000). None of the ML DSM maps proved more accurate than the OK maps on field scale in this case, meaning that legacy soil and yield data could not be used on field scale to create comparable soil property maps for use on field scale PA.