Out-of-fold-guided risk-controlled output calibration for soil organic matter determination using FT-NIR spectroscopy

Fourier-transform near-infrared (FT-NIR) spectroscopy enables rapid and non-destructive determination of soil organic matter (SOM), but small calibration datasets may produce prediction-range compression and concentration-dependent residual bias relative to reference SOM values. This study presents a safe adaptive output calibration convolutional neural network (SAOC-CNN), an out-of-fold (OOF)-guided framework that evaluates conditional output correction and retains the uncorrected base output when the training-only error evidence is insufficient. In second-cross-fitted output-stage validation of the pre-generated CNN predictions on 496 training samples, base was selected in all five outer folds. Quantile mapping with residual refinement (QMap-RR) improved the robust range ratios R₉₀ from 0.7858 to 0.9202 and R₅₀ from 0.7782 to 0.9349, but its RMSE was higher than base (5.2707 versus 5.0950 g kg^-1). On the pre-specified 124-sample within-study holdout, SAOC-CNN increased R² from 0.8248 to 0.8525 and reduced RMSE from 4.5559 to 4.1814 g kg^-1 and MAE from 3.4870 to 3.1785 g kg^-1, corresponding to an 8.22% RMSE reduction. The 330 × 6 configuration matrix was interpreted only as a correlated descriptive analysis. These results support SAOC-CNN as a risk-controlled output-calibration framework and demonstrate a substantial QMap-RR gain under the pre-specified holdout, but do not establish universal superiority across resampling schemes or geographical settings.