Prediction of the concentration of antimony in agricultural soil using data fusion, terrain attributes combined with regression kriging

Abstract

Potentially toxic elements in agricultural soils are primarily derived from anthropogenic and geogenic sources. This study aims to predict and map antimony (Sb) concentration in soil using multiple regression kriging in two distinct modeling approaches, namely Sb prediction using data fusion coupled with regression kriging (scenario 1) and Sb prediction using data fusion, terrain attributes, and regression kriging (scenario 2). Cubist regression kriging (cubist_RK), conditional inference forest regression kriging (CIF_RK), extreme gradient boosting regression kriging (EGB_RK) and random forest regression kriging (RF_RK) were the modeling techniques used in the estimation of Sb concentration in agricultural soil. The validation results suggested that in scenario 1, EGB_RK was the optimal modeling approach for Sb prediction in agricultural soil with root mean square error (RMSE) = 1.31 and mean absolute error (MAE) = 0.61, bias = 0.37, and high coefficient of determination R2 = 0.81. Similarly, the EGB_RK was also the optimal modeling approach in scenario 2, with the highest R2 = 0.76, RMSE = 0.90, bias = 0.06, and MAE = 0.48 values than the other regression kriging modeling approaches. The cumulative assessment suggested that the EGB_RK in scenario 2 yielded optimal results compared to the respective modeling approach in scenario 1. The uncertainty propagated by the modeling approaches in both scenarios indicated that the degree of uncertainty during the modeling process was distributed across the study area from a low to a moderate uncertainty level. However, cubist_RK in scenario 2 exhibited some elevated spots of uncertainty levels. As a result, the combination of data fusion, terrain attributes, and regression kriging modeling approaches produces optimal results with a high R2 value, minimal errors as well as bias. Furthermore, combining terrain attributes with data fusion is promising for reducing model error, bias and yielding high-accuracy predictions.