A General Prediction Model for Organic Matter at Different Soil Depths in the Urumqi River Basin

ABSTRACT Soil organic matter (SOM) constitutes a critical component of terrestrial carbon pools, yet accurate quantification across soil depths remains challenging in arid basins. While hyperspectral techniques enable efficient SOM prediction, current models are limited by: (1) predominant focus on topsoil or single land‐use types; (2) spectral interference from multifactorial coupling effects; (3) insufficient generalization across heterogeneous environments. To address these gaps, we collected 118 multi‐depth soil samples from the Urumqi River Basin and measured their visible‐near‐infrared spectra using a portable spectrometer. Additionally, five spectral preprocessing techniques were employed, including Savitzky–Golay smoothing, Multiplicative Scatter Correction, Standard Normal Variate, first derivative, and second derivative, along with three feature selection methods (Boruta, Successive Projections, and Uninformative Variable Elimination). A suite of predictive algorithms, including Partial Least Squares Regression (PLSR), Support Vector Machines (SVM), Multiple Linear Regression (MLR), and Random Forest (RF) classifiers, were systematically evaluated to develop a generalized model for predicting SOM content at different soil depths. These multivariate statistical approaches leverage cross‐sectional data to enhance the robustness of SOM prediction through non‐invasive spectral data integration. The findings indicate: (1) The second derivative transformation exhibited prominent advantages in spectral pretreatment by amplifying subtle absorption features obscured in raw spectra. Through effective suppression of scattering interference and baseline drift, second derivative (SD)‐processed data substantially improved feature extraction efficiency, thereby elevating the robustness and generalization capacity of subsequent quantitative models. (2) All feature selection methods improved model predictive accuracy, with the Boruta algorithm outperforming SPA and UVE based on PLSR, SVM, and RF models. (3) The comparative analysis of modeling approaches revealed substantial differences in prediction accuracy for SOM across the study area. Among the evaluated models, the SD‐Boruta‐RF algorithm demonstrated superior performance, achieving notable validation metrics with a coefficient of determination ( R 2 ) of 0.89, accompanied by RMSE and MAE values of 28.24 and 15.47 g/kg, respectively. This study confirms that the SD‐Boruta‐RF framework demonstrates strong predictive capabilities for modeling organic matter at different soil depths in the Urumqi River Basin. The developed methodology provides an effective technical solution for continuous organic matter monitoring and spatial estimation in arid ecosystems.