A modified radial basis function network integrating Mahalanobis distance and LASSO for soil mapping

Abstract

In recent decades, machine learning (ML) has replaced geostatistics in soil mapping to overcome the lack of regional soil data. The radial basis function network (RBFN) is a classical ML method for spatial interpolation, but it has limitations in terms of anisotropy and flexibility in soil science. To overcome these shortcomings, we proposed a new model called RBFN-S by incorporating the Mahalanobis distance and the least absolute shrinkage and selection operator (LASSO) into the conventional RFBN, which are new developments in spatial interpolation. One simulated and two real cases were used for validation and investigation. The results showed that RFBN-S performed well in interpolating both Gaussian fields and non-stationary and non-Gaussian fields. In addition, the two introduced elements also played a significant role in improving the interpolation. The comparisons between RBFN-S and conventional RBFN, LASSO, and geostatistical methods demonstrate the advantage of RBFN-S in capturing correlations between soil properties and their location. Further development of the RBFN-S can be performed from the perspective of uncertainty quantification, optimization of meta-parameter tuning, and incorporation of environmental covariates.