Comparing Visible/NIR and MIR Hyperspectrometry for Measuring Soil Physical Properties

Abstract. Rapid developments in semiconductor technologies has improved the robustness and reduced the cost, size and complexity of hyper-spectral instruments, hence, making them suitable for deployment in the field for in situ determination of soil physical and chemical attributes. The goal of this research was to compare two portable, field-deployable hyperspectral sensor systems. The first was a portable mid-infrared (MIR) spectrometer operating between 5,500 and 11,000 nm. The second was a dual spectrometer system operating in both visible (Vis) and near-infrared (NIR) (400‑2,200 nm). A large archived set of 282 soil samples (collected on fields from four Canadian provinces) was used to represent an extensive range in soil textures, varying from sand to clay loam soils with a substantial range of soil organic carbon (SOC). Conventional soil analyses were performed on these samples prior to the measurements by both instruments. Wherever needed, the spectral data were transformed into optical density measurements and then pre-treated by mean centering (MC). Both data sets were randomly partitioned into training (70%) and validation (30%) sets. Partial least squares regression (PLSR) was used to develop spectral calibrations for predicting the percentage of sand, clay and SOC. Both Vis-NIR and MIR spectra revealed similar results. However, clay was better predicted using Vis-NIR and sand using MIR spectra. The highest coefficients of determination (R2) were found for sand (0.82) and clay (0.82). The corresponding root mean squared error (RMSE) was 10% and 7%, respectively. The ability to accurately predict SOC was not as well supported for the set of soils used in this experiment with a root mean squared error of approximately 0.4%. The tested methods prove the usefulness of both portable Vis-NIR and MIR spectrometers for predicting soil texture. Neither sensor showed substantially better performance.