Portable near-infrared spectroscopy with wavelength selection and machine learning for efficient prediction of multiple soil properties.

Portable near-infrared (NIR) spectroscopy represents a rapid, non-destructive, and cost-effective alternative to conventional laboratory methods for soil analysis. This study integrates portable NIR spectroscopy with spectral pre-processing, wavelength selection, and machine learning (ML) techniques to estimate ten soil properties: bulk density, calcium carbonate, clay, extractable potassium, organic carbon, sand, silt, total carbon, total nitrogen, and soil pH. Two methodological approaches were assessed. The first approach involved pre-processing raw spectra followed by Principal Component Analysis (PCA). The second approach utilized Extra Trees Regressor (ETR)-based wavelength selection to identify the 150 most informative wavelengths. The ETR-selected wavelengths substantially improved five-fold cross-validation performance for calcium carbonate (R2 = 0.949 ± 0.015), organic carbon (R2 = 0.924 ± 0.015), total carbon (R2 = 0.919 ± 0.026), total nitrogen (R2 = 0.893 ± 0.015), and soil pH (R2 = 0.884 ± 0.005). Statistical analyses demonstrated that the acquired spectra were consistent and the proposed framework delivered robust and reliable predictive performance across the five-fold cross-validation experiments.