Soil moisture effects on predictive VNIR and MIR modeling of soil organic carbon and clay content

Abstract

Portable visible to near-infrared (VNIR) and mid-infrared (MIR) soil spectroscopy holds great potential to support field applications in soil science and management by complementing conventional soil analytical methods. Under field conditions, however, soil moisture can critically affect the quality of reflectance measurements. In this study, we examined the effects of soil moisture on VNIR and MIR soil spectra and how its magnitude and variation impact the accuracy and robustness of predictive spectral models. We carried out a systematic re-wetting experiment on two soil datasets of different scale and origin that were measured at four gravimetric moisture levels (air-dried, 5 %, 10 %, 15 %) with portable VNIR and MIR instruments. The spectral data of each moisture class, as well as randomized combinations of different moisture contents, were then used to calibrate VNIR, MIR and combined PLSR models to estimate soil organic carbon (SOC) and clay content, where combined models included spectra concatenation (VNMIR) and model output average (MOA). The overall shape of MIR spectra was more significantly distorted by soil moisture than VNIR spectra, while the general impact of soil water content in both spectral domains was texture-dependent. In terms of predictive accuracy, MIR models were generally superior for air-dried sample material, while VNIR models fared better for uniformly moist samples. With increasing soil moisture variability, comparative estimation accuracies between individual VNIR and MIR models were dependent on the underlying dataset. VNMIR and MOA models proved beneficial and yielded the most accurate and robust predictions for SOC and clay content when soil moisture was variable, irrespective of the considered dataset (regional dataset: RMSESOC = 0.22–0.27 %, RMSECLAY = 2.67–3.14 %; field dataset: RMSESOC = 0.09–0.11 %, RMSECLAY = 0.87–1.15 %). Predictive mechanisms, as evaluated by variable importance in the projection (VIP) of PLSR models, changed substantially with variation in soil water content, especially in the MIR, where important absorption bands for SOC and clay minerals could be heavily attenuated or completely masked. Our study highlights the advantages of employing both VNIR and MIR instruments for spectral data collection on soils in field condition and the potential of integrating VNIR and MIR spectra collected at different soil moisture levels into soil spectral libraries.