Combining Spectral Preprocessing With Joint Wavelength Selection to Improve the Accuracy of Carbon and Nitrogen Assessments in Farmland Soils in the Loess Plateau Area of China

Soil organic carbon (SOC), soil total nitrogen (STN), and soil carbon-to-nitrogen ratio (C/N) are crucial indicators for assessing soil quality and health, as well as key parameters reflecting greenhouse gas emissions and climate change processes. However, the rapid and accurate estimation of SOC, STN, and C/N based on near-infrared spectroscopy still faces challenges such as wavelength redundancy and low modeling accuracy. This study takes the farmland soil in central and southern Shanxi Province as the object and constructs the SOC, STN, and C/N inversion processes covering spectral pretreatment, variable selection, and multimodel modeling. The impact of 11 pretreatment methods and their combinations on the correlation of spectral-soil attributes was systematically evaluated, and the combined variable selection method was innovatively introduced to improve the stability and interpretability of feature extraction, and combined models such as support vector machine (SVM), partial least squares (PLS), and multiple linear regression (MLR) were combined to model and compare and analyze the full-spectrum and selected bands. The results show that the variable importance in the projection-successive projections algorithm (VIP-SPA) wavelength selection method has more advantages than a single SPA algorithm and can significantly extract the key band intervals of SOC, STN, and C/N to enhance the model interpretability. In terms of spectral pretreatment, the use of Savitzky–Golay smoothing combined with firstorder derivatives can effectively improve the correlation between the spectrum and soil properties and improve modeling effect. Among them, the optimal model of SOC is the SVM model based on this preprocessing method (R2v=0.7258, RMSEv=2.5343, RPDv=1.8770), while the PLS model based on Savitzky–Golay smoothing combined with first-order derivative preprocessing performed best (R2v=0.6428, RMSEv=0.5161, RPDv=1.6710), while the optimal estimation model of C/N is the SVM model based on first-order derivative (R2v=0.7479, RMSEv=3.2608, RPDv=1.9956). In addition, the VIP-SPA-MLR combination model is better than SPA-MLR in various indicators, verifying the advantages of multi-algorithm fusion in improving model robustness. This study combines a variety of spectral optimization and modeling strategies to significantly improve the estimation accuracy of farmland soil SOC, STN, and C/N, expands the application potential of near-infrared spectroscopy technology in rapid soil diagnosis and sustainable agricultural management, and has important theoretical value and practical significance