Quantifying the added value of remote-sensing covariates in soil pedotransfer functions: a Bayesian hierarchical assessment with permutation controls for Mediterranean croplands in Cyprus

Accurate estimates of soil nutrients and organic carbon are essential for sustainable land management, yet the factors that most improve soil pedotransfer function (PTF) accuracy remain poorly quantified. Remote sensing (RS) is increasingly used to augment soil databases, but its benefits are inconsistent and a genuine RS signal is rarely separated from accuracy gained by adding predictors. This study framed PTF development as three modelling decisions: algorithm choice, feature selection, and whether to augment soil predictors with an RS group. Their effects on predictive error for nitrogen (N), phosphorus (P), potassium (K), and organic carbon (OC) were quantified across 951 agricultural sites in Cyprus using a Bayesian hierarchical model with a permutation-based pseudo-RS control that isolates genuine RS signal from added dimensionality. Model selection had the greatest impact, with ensemble and tree-based models reducing RMSE by up to -41.4%. Feature selection ranked second, with wrapper methods (FRFS, RFE) reducing OC RMSE by about -10%. Augmenting predictors with the RS group yielded smaller, property-dependent net gains; permutation controls confirmed a genuine but consistently smaller RS signal, consistent with the mismatch between coarse RS support and point-scale soil measurements. This does not imply that remote sensing is of limited value in general; rather, in this information-rich laboratory dataset, RS covariates added relatively little on top of the measured soil properties. For point-based PTFs in Mediterranean croplands, algorithm and predictor selection yield larger, more reliable returns than adding RS covariates. It provides a reusable framework that quantifies uncertainty in comparisons among modelling decisions.