Assessment of Reflectance Metrics and PLSR Modeling for Soil Organic Matter Prediction Using VIS–NIR–SWIR Spectra

Visible–near infrared–shortwave infrared (VIS–NIR–SWIR, 350–2500 nm) soil reflectance spectroscopy provides a rapid and non-destructive approach for characterizing soil properties, yet the relative contribution of spectral integration metrics, preprocessing strategies, and diagnostic wavelength regions remains an active area of research. In this study, a dataset of 220 soil samples from 30 agricultural sites representing contrasting pedological conditions was analyzed using laboratory reflectance spectra acquired over the full 350–2500 nm range. In addition to point reflectance values, two cumulative reflectance variables were computed: a visible cumulative reflectance (VCR, 350–780 nm) and a full-spectrum cumulative reflectance (CR, 350–2500 nm), designed to capture integrated spectral behavior related to soil color and overall albedo. Partial least squares regression (PLSR) models were developed to predict soil organic matter determined by wet oxidation (SOM), and the influence of spectral range and preprocessing was systematically evaluated.

In particular, VCR emerged as a particularly meaningful descriptor of soil optical behavior, functioning analogously to an integrated soil color or saturation metric and revealing distinct linear trends associated with different soil groups.

The PLSR models based on the full VIS–NIR–SWIR range consistently outperformed those restricted to the visible domain. The most robust and parsimonious configuration used original reflectance data subjected to mean centering, standard normal variate correction, and detrending. Under this configuration, SOM was predicted with coefficients of determination of approximately R² ≈ 0.58–0.61 and RMSE values near 0.8, using 6–7 latent variables. More aggressive preprocessing strategies, including second derivatives and extensive Savitzky–Golay smoothing, produced only marginal improvements while increasing the spurious effects of noise or reducing interpretability, and were therefore deemed unnecessary.

The Variable Importance in the Projection (VIP) trace revealed that SOM prediction was primarily controlled by a limited number of SWIR absorption regions centered near 552, 1414, 1918–2008, 2140, and 2201–2216 nm, consistent with absorptions associated with organic matter and clay minerals—mainly montmorillonites and kaolinite—with comparatively lower influence from calcite and oxides.

Overall, the results demonstrate that VIS–NIR–SWIR spectroscopy up to 2500 nm enables a physically interpretable estimation of soil organic matter with predictive performance across very different soil types.However, the results also show that SOM prediction remains highly significant when PLSR models are built exclusively from visible-range data (380–780 nm). The findings highlight the utility of cumulative reflectance metrics and the importance of prioritizing model robustness and spectral interpretability over excessive spectral manipulation, supporting the application of full-range soil spectroscopy for soil characterization and mapping at larger scales.

 

Acknowledgements, this research was funded by EJP-SOIL grant agreement 862695. Javier González Canales received funding through grant PRE2021-097966 from MCIU/AEI/10.13039/501100011033 and the European Social Fund (ESF, Investing in Your Future)