Inversion model of soil organic matter content in black soil by Vis-NIR spectroscopy using adaptive stacking

Accurate prediction of soil organic matter (SOM) is crucial for precision agriculture in Northeast China’s black soils (Chernozems, WRB 2022). This study proposes a novel Stacking ensemble learning framework using Visible and Near-Infrared (Vis-NIR) spectroscopy, integrating First Derivative (FD) and Multiplicative Scatter Correction (MSC) preprocessing, Competitive Adaptive Reweighted Sampling (CARS), and XGBoost feature selection with a dynamic weight allocation strategy for SOM ranges (2.55–31.05 g/kg, split at >20 g/kg and <20 g/kg). Combining Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and XGBoost, the FD+CARS+Stacking model achieved superior accuracy (R²=0.8364, RMSE=2.2472) compared to traditional models (R²p=0.50–0.80). Pearson’s correlation analysis validated preprocessing and feature selection efficacy, with FD enhancing correlations at 523 nm (r=-0.7345, p<0.01), MSC at 2221 nm (r=0.7502, p<0.01), and CARS and XGBoost retaining high-correlation bands (e.g., 523 nm for FD+CARS, 2223 nm for MSC+XGBoost). Based on 291 soil samples from Nong’an County, Jilin Province, China, this framework provides a robust tool for high-precision SOM estimation, supporting soil health monitoring and optimizing nutrient management in black soil regions. • Enhanced Vis-NIR spectroscopy for accurate black SOM estimation. • Combines preprocessing and feature selection to enhance SOM model analysis. • Compares single and Adaptive Stacking models for SOM inversion applicability. • A new approach improves SOM estimation in agricultural fields.