Assessing the robustness of VIS-NIR spectroscopy-based soil organic carbon prediction against four wet chemistry methods

A visible and near-infrared (VIS-NIR) spectroscopy-based machine learning model can rapidly predict soil organic carbon (SOC) while minimizing environmental impacts once established. Nevertheless, model reliability has mostly been evaluated statistically, with limited explanation from a soil science perspective or validation against various wet lab methods, which differ in reliable. Therefore, this study assessed the robustness of a VIS-NIR-based Cubist model trained with Walkley-Black (WB) SOC data (<i>n</i> = 400) from diverse land uses and soil types in Taiwan. Additionally, SOC contents by total organic carbon (TOC) analyzer, elemental analyzer (EA), and loss-on-ignition (LOI) methods from an independent sample set (<i>n</i> = 46) tested the model’s applicability. The model achieved R2=0.72 and RMSE=1.09%, whilst highly aligned with WB measurements (LCCC=0.82). Major model predictors indicated SOC with varying stabilities, where degradation was linked to weathering and stabilization to iron oxides. Predicted SOC contents also aligned with TOC (LCCC = 0.81) and EA (LCCC = 0.76), but less with LOI (LCCC = 0.48) due to overestimation from structural water loss. While VIS-NIR model reflected the precision of training data, consistency with other methods supported its utility as a complementary tool for SOC estimation. This study highlighted VIS-NIR spectroscopy’s potential in evaluating stabilized carbon pool for carbon monitoring.