Eliminating the influence of free iron oxides on the prediction of organic matter in red soils using Vis-NIR reflectance spectroscopy

Free iron oxides (Fe d ) in red soils negatively affect the accuracy of organic matter (OM) prediction using visible and near-infrared (390–1000 nm) spectroscopy. Based on a typical red soil sample set (n = 36, including both Fe d - and OM-removed samples), this study investigated the spectral characteristics of Fe d and OM, and applied the maximum overlapping discrete wavelet transform (MODWT) to extract wavelengths associated with Fe d spectral features across different wavelet components. We evaluated whether removing the extracted Fe d spectral features from the red soil spectra could improve the accuracy of OM prediction by using three datasets: (A) regional-scale, laboratory-measured dry soil spectral dataset; (B) field-scale, laboratory-measured moist soil spectral dataset; and (C) field-scale, unmanned aerial vehicle-measured soil spectral imaging dataset. Three calibration algorithms—partial least squares regression (PLSR), linear support vector machine (LSVM), and random forest (RF)—were employed for OM prediction. Compared with raw spectra as a baseline, preprocessing via Fe d spectral feature removal improved the prediction accuracy of OM in red soils for all datasets and algorithms tested. The average percentage improvement across all datasets is 19.63 % in RMSE val-ori and 93.80 % (absolute values increased by 0.21 averagely) in R 2 val-ori . The results confirm that wavelet-decomposed Fe d spectral feature removal preprocessing improves the accuracy of OM spectral prediction in red soils, and demonstrate the transferability of the feature extraction parameters across different application scenarios.