Fuzzy cluster membership as a reliability indicator for continental soil spectroscopy models

Visible–near-infrared soil spectroscopy has matured into a reliable solution for rapid, non-destructive, and cost-efficient estimation of multiple soil properties from a single measurement. At continental scale, however, large soil spectral libraries are inherently heterogeneous: spectra–property relationships vary across soil-forming conditions and land-use settings, which can compromise the validity of a single global calibration and produce geographically or environmentally structured sources of prediction error that are not always apparent from standard model evaluation metrics. This work proposes a reliability-aware, cluster-informed modeling framework for pan-European soil spectral libraries, organizing the library into coherent domains using ancillary descriptors (geographic location and broad environmental stratification) solely to support library structuring and prediction routing; all regression models remain strictly spectra-driven. Using the LUCAS 2015 topsoil dataset, four target properties (soil organic carbon, total nitrogen, pH, and calcium carbonate) are evaluated under a global baseline, hard cluster routing, and two fuzzy aggregation variants. A non-linear ensemble learner (random forest) and a linear chemometric baseline (partial least squares regression, PLSR) are benchmarked to assess model-dependence of routing benefits. Across properties, cluster-informed routing produced the clearest gains for PLSR. Hard routing reduced CaCO 3 RMSE from 32.53 to 22.53 g k g − 1 and pH RMSE from 0.61 to 0.47, while increasing the corresponding R 2 values from 0.82 to 0.91 and from 0.75 to 0.85, respectively. For RF, aggregate improvements were more limited, although Hard routing produced lower absolute error than the Global model for 63.06% of CaCO 3 test samples. Unconstrained fuzzy averaging across all experts was consistently less effective than selective routing, with particularly clear degradation for RF, whereas restricting aggregation to dominant memberships substantially mitigated this loss and, for PLSR, largely preserved the benefits of cluster-specific prediction. Maximum fuzzy membership demonstrated a link with routing reliability, most clearly for RF: higher-membership samples generally exhibited lower rates of degradation for selected properties, although this relationship was not uniform across properties or model families. Overall, the results support cluster-informed modeling as a complementary strategy to global calibration that enhances reliability assessment.