Machine Learning and Geospatial Modelling of Soil Zinc: Implications for Crop Nutrition in the Brazilian Cerrado

Ensuring sustainable crop nutrition requires improving Zn phytoavailability in weathered and leached tropical soils, which depends not only on available Zn but also on soil pH, texture and P levels. This study developed a digital soil mapping framework for the 200‐million‐hectare Cerrado using 32,272 topsoil (0–20 cm) samples of available Zn, 42,163 of available P, and 42,417 for pH, mostly from precision‐management systems, applying spatial machine learning models for Zn and key soil properties (pH, texture and available P) and formalizing Zn deficiency conditionals to map Zn phytoavailability. Higher accuracy was obtained for Zn, followed by P and pH, with RFE–random forest models achieving R 2 values of 0.61 for Zn, 0.41 for P and 0.43 for pH, with corresponding RMSEs of 1.65, 5.75 mg dm −3 and 0.37. All predictions showed low uncertainty. Both conventional and newly developed environmental covariates, designed to capture anthropogenic influence, were ranked among the most important predictors. Overall, 78.3% of the Cerrado showed a medium Zn deficiency risk. The highest deficiency occurred under Zn‐pH interaction, affecting 4% (~7.7 million ha), followed by Zn‐Clay (3%, ~5 million ha). Medium risk was most pronounced for Zn‐pH (77%, ~158 million ha) and Zn‐P rainfed (43.6%, ~88.5 million ha). Low risk was observed in 91%, 89%, and 55% of the area for Clay, P irrigated and P rainfed interactions, respectively. Accounting for soil properties such as pH, clay, and P rather than relying solely on extractable Zn is essential for accurately diagnosing Zn deficiency and estimating fertilizer needs. Zn demand across high‐risk areas was estimated at ~38.5 billion kg when considering other soil factors, compared with only ~7 billion kg when based solely on Mehlich‐1 Zn levels, a difference more than fivefold.