C G Kusuma, J Saralakumari, S. Dharumarajan, A. Sathish, N. Umashankar
Accurate estimation of soil organic carbon (SOC) is essential for soil quality assessment, carbon management and sustainable agriculture, particularly in semi-arid regions characterised by high spatial variability. Visible and near infrared (Vis–NIR) spectroscopy has emerged as a rapid and non-destructive technique for SOC prediction; however, its performance is strongly influenced by spectral preprocessing and modelling approaches. Despite growing interest, systematic evaluation of preprocessing-model interactions for SOC prediction under semi-arid conditions in India remains limited. This study aimed to systematically evaluate the effect of different preprocessing techniques and modelling methods on SOC prediction using Vis-NIR spectral data from a semi-arid region of southern India. A total of 300 surface soil samples were collected and analysed for SOC using the Walkley–Black method. Spectral data (400–2400 nm) were preprocessed using raw spectra, Savitzky–Golay first and second derivatives (SG-1D and SG-2D), multiplicative scatter correction (MSC) and standard normal variate (SNV). Partial least squares regression (PLSR) and random forest (RF) models were developed using calibration (75 %) and validation (25 %) datasets. Results showed that spectral preprocessing significantly influenced model performance, with SG-1D and MSC providing consistent improvements by enhancing spectral features and reducing scattering effects. The RF model outperformed PLSR, with the best performance achieved using RF combined with MSC and SG-1D (validation R2 = 0.50, RMSE = 0.27 %, RPIQ ≈ 1.66–1.67). The findings indicate that optimal SOC prediction depends on the combined selection of preprocessing techniques and modelling approaches. This study provides a practical and reliable framework for improving soil spectroscopy applications in digital soil mapping and precision agriculture.