Modeling Soil pH at regional scale using environmental covariates and machine learning algorithm

Soil pH serves as a critical indicator of soil chemistry and fertility, and mapping its spatial distribution holds significant importance for effective crop management. Digital soil mapping (DSM) is a commonly employed method for making rapid and cost-effective quantitative predictions of soil properties and soil classes. In the present study, we mapped soil pH (0–15 cm) on a regional scale in Karnataka using a combination of various environmental variables. Three distinct machine learning models, namely support vector machine (SVM), Cubist, and random forest (RF), were assessed using a dataset of 146,044 observations collected under various projects. The environmental covariates used for soil pH prediction encompassed terrain attributes, Landsat-8 data, vegetation indices, and climatic variables. Among these models, RF model exhibited the most acceptable results for predicting soil pH (R²_val = 0.61, CCC_val = 0.74, RMSE_val = 0.66). On the other hand, the Cubist and SVM models displayed comparatively lower accuracy, explaining only about 46–49% of the variation. The inclusion of climatic variables and Landsat-8 data emerged as crucial factors for predicting soil pH. The study successfully produced high-resolution maps of soil pH for the entire state at a 90-m resolution, while also quantifying the associated uncertainty. These high-resolution maps have the potential to be valuable for decision-makers, stakeholders, and agricultural practitioners towards precision agriculture and land resource management.