Evaluation of error structure in near-infrared spectroscopy modeling of soil based on error covariance and...

Near-infrared spectroscopy offers a rapid and non-destructive approach for soil moisture content (SMC) prediction, yet its accuracy is constrained by heterogeneous error structures arising from soil complexity, light scattering, and instrumental noise. Current preprocessing strategies predominantly rely on empirical trial-and-error, lacking systematic analysis of error sources and their mitigation mechanisms. This study proposes an error structure-guided framework integrating error covariance matrix (ECM) and correlation matrix analysis to quantify heteroscedasticity and error coupling in raw soil spectra. Identifying dominant error types (e.g., multiplicative noise at OH absorption peaks, 1450 nm and 1940 nm) and their origins by ECM, we optimized preprocessing selection and feature band extraction. Experimental results demonstrated that multiplicative scatter correction (MSC) and standard normal variate (SNV) effectively addressed baseline shifts and heteroscedastic errors, reducing average error correlations from 0.99 to 0.20. Combined with competitive adaptive reweighted sampling (CARS), the PLS model achieved superior performance (testing set R² = 0.99, RPD = 9.1). The proposed framework provides a universal strategy for error-aware spectral modeling, extendable to multi-parameter soil analysis (e.g., organic carbon, pH), and offers technical support for field-deployable NIR systems in precision agriculture. Future research would systematically evaluate the applicability of the framework across diverse soil matrices, including but not limited to clay and sandy loam.