Using geostatistics and machine learning models to analyze the influence of soil nutrients and terrain attributes on lead prediction in forest soils

The study aimed at investigating the possibility of predicting lead (Pb) in forest soils by combining terrain attributes and soil nutrients using geostatistics and machine learning algorithms (MLAs). The study was partitioned into three categories: predicting Pb in forest soil using terrain attributes and RK (Context 1); predicting Pb in forest soil using soil nutrients and RK (Context 2); and lastly predicting Pb in forest soils using a combination of soil nutrients, terrain attributes, and RK (Context 3). Stochastic Gradient Boosting-regression kriging (SGB-RK), cubist regression kriging (CUB_RK), quantile regression forest kriging(QRF_RK) and k nearest neighbour regression kriging (KNN_RK) were the modeling approaches used in the estimation of lead (Pb) concentration in forest soil. The results showed that combining the terrain attribute as an auxiliary dataset with QRF_RK proved to be the most effective method for predicting Pb in forest soil (context 1). The most effective method for predicting Pb in forest soil was to combine soil nutrients as an auxiliary dataset with SGB_RK (context 2). Combining cubist_RK with an ancillary dataset of soil nutrients and terrain attributes is the most effective method for predicting Pb in forest soils (context 3). In addition, combining ancillary variables such as soil nutrients and terrain attributes with cubist_RK generated the best results for estimating Pb concentration in forest soils. It was found that applying a robust digital soil mapping (DSM) model in combination with terrain attributes and soil nutrients is efficient in predicting the spatial distribution and estimation of uncertainty levels of lead (Pb) in forest soils based on the model’s accuracy parameters.