Spatial prediction of soil micronutrients using machine learning algorithms integrated with multiple digital covariates

The design and application of multiple tools to spatially assess soil properties are key to efficient land management plans. While soil micronutrients are paramount soil properties for multiple purposes, collecting a representative number of samples is time-consuming and expensive. The main aim of this research was to evaluate different scenarios involving 41 environmental variables with two machine learning (ML) algorithms for predicting the spatial distribution of soil micronutrients within the piedmont plain in north-eastern Iran. Sixty-eight locations with different land uses were soil sampled to determine the contents of the micronutrients of iron (Fe), manganese (Mn), zinc (Zn) and copper (Cu). The environmental variables were derived from a digital elevation model, open-source Landsat 8 OLI, Sentinel 2A MSI images, WorldClim climate variables and raster maps of key soil properties. Normalised Root Mean Square Error (NRMSE) and Taylor diagrams were used to evaluate the ML models. Based on the validation results and soil scientists evaluation of the produced maps, the Random Forest (RF) algorithm emerged as the most effective method for predicting the spatial distribution of the soil micronutrients. For the validation set, 91%, 94%, 91% and 108% NRMSE values for Fe, Mn, Zn and Cu, respectively, were given by the RF algorithm. However, one parsimonious scenario involving only the climate covariates also showed promising results. These digital maps produced at 30 m spatial resolution could be used as valuable reconnaissance base information to effectively identify micronutrient deficiencies and excess hotspots for large areas.