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

The design and application of multiple tools to map soil micronutrients is key to efficient land management. While collecting a representative number of soil samples is time consuming and expensive, digital soil mapping could provide maps of soil properties fast and reliably. The objective of this research was to predict the spatial distribution of soil micronutrients within the piedmont plain in northeastern Iran using random forest (RF) and support vector regression (SVR) algorithms. Sixty-eight locations with different land uses were sampled to determine the content of iron, manganese, zinc and copper in the topsoil (0–20 cm). Forty-one digital covariates were used as input to the models and were derived from a digital elevation model, open-source remote sensing (RS) data (Landsat 8 OLI and Sentinel 2A MSI images), WorldClim climate database and maps of soil properties. Covariates were grouped into 11 scenarios: I–III, based on RS data; IV–VI, including RS, topographic, climate and soil covariates; VII, VIII and IX, based only on topographic, climate and soil covariates, respectively; X and XI, based on recursive feature elimination and expert opinion, respectively. The RF algorithm gave 91, 94, 91 and 108% normalized root mean squared error values for iron, manganese, zinc and copper, respectively, for the validation dataset with scenario XI. The most important digital covariates for micronutrients prediction with both RF and SVR models were precipitation seasonality, mean annual temperature and the mean saturation index based on Sentinel 2A MSI data. Digital maps produced at 30 m spatial resolution using scenario XI could be used to effectively identify micronutrient deficiencies and excess hotspots.