Sample-Level Validation and Applicability of Near-Infrared Spectroscopy for Rapid Screening of N, P2O5, and K2O in Compound Fertilizers

Nitrogen (N), phosphorus pentoxide (P2O5), and potassium oxide (K2O) are key quality indicators of compound fertilizers. This study evaluated the sample-level predictive performance and applicability of compact near-infrared (NIR) spectroscopy for rapid screening of these nutrients in unground granular compound fertilizers. A total of 143 samples were measured using two independent loadings per sample, yielding 286 spectra over 900–1700 nm. Candidate modeling strategies were compared using sample-ID-grouped repeated nested cross-validation. Model stability was examined by loading-repeatability sensitivity analysis and leave-one-acquisition-date-out validation, while nutrient covariation was assessed using nutrient-only linear reference models. The selected procedures were analyte-specific ensemble models for N, P2O5, and K2O. Internal coefficients of determination (R2) were 0.9005, 0.8279, and 0.8034, with root-mean-square error (RMSE) values of 0.8392, 1.0884, and 1.5724 percentage points, respectively.. Relative to nutrient-only linear reference models, RMSE decreased by 32.7% for P2O5 and 31.9% for K2O. Without refitting, the finalized models were applied to 29 subsequent production-batch samples from the same source, yielding R2 values of 0.7147, 0.7496, and −2.2911 for N, P2O5, and K2O, respectively. N and P2O5 retained potential for rapid preliminary screening under the evaluated conditions, whereas the K2O relationship did not transfer reliably, indicating the need for broader calibration coverage across production batches and formulations.