Clay content mapping and uncertainty estimation using weighted model averaging

Abstract Accurate prediction of clay is the basis for soil quality assessment and decision making in land use because it governs soil moisture and fertility dynamics. However, using laboratory methods to determine clay across a large district and at multiple depths is tedious and expensive. An alternative is to use proximally and remotely sensed digital data, that can be coupled to laboratory measured clay through models. This study aims to predict topsoil (0–0.3 m) and subsoil (0.9–1.2 m) clay at district scale by comparing; i) importance of proximally (i.e., apparent soil electrical conductivity – ECa) and remotely (i.e., γ-ray spectrometry, digital elevation model – DEM) sensed data, ii) models including a linear mixed model (LMM) and machine learning models (MLs, i.e., Cubist, random forest [RF], support vector machine regression [SVMR], quantile regression forests [QRF], extreme gradient boosting [XGBoost] and bagEarth), iii) two model averaging techniques (i.e., Granger–Ramanathan averaging (GRA) and Lin’s concordance (LCCC) weights) from the top four best models, and iv) uncertainty of the prediction. The results showed that the γ-ray data was most important for topsoil clay prediction, while in the subsoil the slope was most important. Moreover, for topsoil clay prediction the RF was best with fair accuracy (RPD = 1.64), followed by QRF (1.62), Cubist (1.61) and LMM (1.55) which outperformed bagEarth (1.51), SVMR (1.47) and XGBoost (1.47). For the subsoil, all seven models achieved poor accuracy (RPD