Applying Machine Learning Techniques to Model and Map Soil Surface Texture Using Limited Legacy Data

Soil texture influences physical, chemical, and biological soil processes and is crucial for understanding soil functions and improving decision making. However, obtaining spatial information on soil texture and its uncertainty is still a challenge, due to high costs on sampling and mapping large areas. We aimed to model and map the surface soil texture in a large and high geodiverse area (Minas Gerais state, Brazil) through machine learning techniques and limited legacy data. We used 667 georeferenced samples containing coarse sand, fine sand, silt, and clay contents at 0–20 cm depth and 109 covariates based on the soil formation factors. Predictor selection and predictive modeling of each particle size for five different models were performed 100 times, using 75% of samples for training and 25% for testing. The Random Forest was the best model for all-grain size distribution. The most important covariates were related to the parent material, soil properties, and climate. We conclude that machine learning techniques can produce fair consistent maps for superficial texture and the associated models’ uncertainties, even using limited legacy data.