Evaluating Different Soil Mapping Methods for Precision Agriculture

Abstract

This study aimed to determine the most accurate and cost-effective method of soil mapping at field scale. To achieve this, a field was selected from the Free State Province in South Africa, and soil analysis was conducted on samples taken on a 1-hectare grid basis. The industry standard method of soil mapping in South Africa, Ordinary Kriging (OK), was evaluated alongside semi???variograms to determine the viability of using 1-hectare grids to sample soils. The accuracy of the OK maps was assessed using 10-fold cross-validation. Furthermore, digital soil mapping (DSM) with Machine Learning (ML) models were used to create maps from the same properties. The two models used were Cubist and Random Forest (RF), while the covariates used included Brightness Index (BI), Redness Index (RI), Saturation Index (SI), Normalised Difference Vegetation Index (NDVI), Digital Elevation Model (DEM) and Harvester Yield Data. The accuracy of the DSM with ML maps was tested by splitting the dataset 70% - 30%, where 70% of the dataset was used to train the model and 30% was used to validate the predictions. The statistics of the OK and DSM with ML maps were compared, and the maps were also visually inspected. Thereafter, fertiliser recommendation maps were created from the potassium (K) and phosphorus (P) maps of both the OK and DSM predictions. From these fertiliser recommendation maps a financial assessment was conducted of the two mapping methods. The semi-variograms showed that the 1-hectare grid system should not be used. The potassium (K) calcium (Ca), magnesium (Mg) and cation exchange capacity (CEC) OK maps all showed directionality. Both the DSM with ML and OK results could not be considered accurate enough. The financial analysis showed that due to the maps being inaccurate, there were huge losses when using both methods of mapping.