Identifying electrical conductivity in topsoil by interpretable machine learning

Implementing computational and statistical intelligence allows understanding variations to make knowledgeable decisions in agriculture and land use. In this study, using the interpretation capacity of functional precision models against the lower-level mechanistic models has been given priority and advanced regression algorithms were used for this purpose. For exploring the relationships between electrical conductivity (EC) which is the most critical indicator for salinity and irrigation and soil parameters (texture, chemical concentrations) a comparative assessment based on supervised learning algorithms has been conducted and the outcomes have been interpreted by statistical learning. Both linear and non-linear statistical approaches have been used to analyze EC since it is a heterogeneous natural feature in topsoil. The implementations showed that the non-linear MARS model has provided the best testing accuracy. In addition to precision, relationships between EC and indicator variables, major modelling components and measure of the parameter effects have been exhibited by estimated probabilities and partial dependence measures. The main drivers for EC assessment in topsoil, specifically pH in water, CaCl2, and nitrogen concentration, have been discovered. The benchmarking results revealed that different from the conventional mechanistic models, interpretable machine learning provides additional interpretation, meta-data and transparency for sustainable soil management and environment.