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

Area of Science:

• Agricultural Science
• Environmental Science
• Geospatial Science

Background:

• Soil pH is a key indicator of soil health and fertility, crucial for effective crop management.
• Digital Soil Mapping (DSM) offers efficient, cost-effective quantitative predictions of soil properties.
• Understanding soil pH spatial distribution is vital for optimizing agricultural practices.

Purpose of the Study:

• To map soil pH (0-15 cm) on a regional scale in Karnataka using digital soil mapping techniques.
• To compare the performance of Support Vector Machine (SVM), Cubist, and Random Forest (RF) models for soil pH prediction.
• To generate high-resolution soil pH maps and quantify prediction uncertainty.

Main Methods:

• Utilized a dataset of 146,044 observations for soil pH prediction.
• Employed environmental covariates including terrain attributes, Landsat-8 data, vegetation indices, and climatic variables.
• Assessed three machine learning models: SVM, Cubist, and RF.

Main Results:

• The Random Forest (RF) model demonstrated superior performance with R²val = 0.61 and CCCval = 0.74.
• Cubist and SVM models showed lower accuracy, explaining only 46-49% of the variation.
• Climatic variables and Landsat-8 data were identified as critical predictors for soil pH.

Conclusions:

• High-resolution (90-m) soil pH maps were successfully generated for Karnataka.
• The developed maps and uncertainty quantification can support precision agriculture and land resource management.
• The study highlights the effectiveness of RF models and specific environmental covariates for regional soil pH mapping.