Preserving and increasing soil organic matter (SOM) has been identified a key strategy for climate change mitigation. Plant residues, fungal and bacterial necromass, and other detritus accumulates in the soil and a portion of this SOM is protected or stabilized from further microbial mineralization. In addition to providing numerous agronomic benefits, it is thought greenhouse gas emissions could be offset through net gains in SOM and the upward climb of global temperatures could be stalled or even reversed. However, storage of atmospheric carbon as SOM is only effective as a climate change mitigation strategy if SOM associated C remains in the soil long-term, i.e. more than 100 years.
Soil organic matter stabilization is known to be controlled by several key mechanisms, i.e. physical occlusion, polyvalent cation bridging, weak interactions like hydrogen bonding and Van der Waals forces, and ligand exchange. These physical and chemical stabilization mechanisms vary in the protection they provide SOM; physical occluded SOM is weakly protected as compared to SOM protected via ligand exchange. While these stabilization mechanisms and the affects they have on SOM have been known for some time, much regarding the soil properties associated with these mechanisms and SOM they stabilize remain a mystery.
Measurements of SOM and other soil properties important for SOM stabilization are necessary for modeling changing soil C levels over time. However, conventional laboratory methods for assessing these soil properties are time consuming and expensive for large numbers of samples. Visible near-infrared spectroscopy (Vis-NIRS) is a rapid method and inexpensive technique for predicting soil properties. Multivariate statistical approaches are used in conjunction with reference soil data representative of the study area to calibrate and validate models, which if a sufficient level of accuracy is achieved can be used to predict unknown samples only with the VNIR spectra. Therefore, the objective of this research is to 1) elucidate which soil properties influence C stabilization in Ohio soils and 2) assess the ability of portable Vis-NIRS to predict SOC and other soil properties thought to impact SOC stabilization using two multivariate statistical calibration approaches: partial least squares regression (PLSR) (linear) and support vector machines (SVM) (non-linear).
Soils evaluated in this study were part of an archival collection of soils sampled from across Ohio as part of the National Cooperative Soil Survey. Soils were sampled by the Soil Survey Staff from 1957-1994 and, after being dried and sieved <2 mm, the Ohio State University (OSU) Soil Characterization Laboratory carried out basic soil characterization to ascertain numerous soil properties (total organic C, total sand, total silt, total clay, fine clay, pH, extractable Ca2+, Mg2+, and K+, cation exchange capacity (CEC), and base saturation). On a subset of soils relative abundance of minerals (illite, smectite, vermiculite, interstratified, interlayer chlorite, kaolinite, and quartz) in the clay-sized fraction were estimated via x-ray diffraction.
Soils were selected from the entire archive collection to represent heterogeneity within each of the four major physiographic regions of Ohio – Till Plains, Huron-Erie Lake Plains, Glaciated Allegheny Plateau and Allegheny Plateau (unglaciated). Soils from each region were selected based on organic C content, clay content and pH in order to capture the full range of these three properties. The final data set included 103 samples from 41 different counties across the state.
In the laboratory, bulk soils were separated into four different SOM fractions hypothesized to represent a continuum of stabilization mechanisms. These fractions, isolated through a physio-chemical-density separation procedure proposed by Zimmermann et al. (2007), are particulate organic matter (POM), sand and stable aggregates (S+A), silt and clay (si + c), and resistant SOC (rSOC). Separated fractions were dried and then ground and homogenized in a roller-mill. Total C and N concentrations of each fraction were measured by elemental analyzer and used to calculate C content (g C kg soil -1), relative % (fraction-content*bulk-content-1*100) and stock (Mg C ha-1). Bulk soils were scanned using diffuse reflectance mid-infrared spectroscopy (MIRS) and specific peak assigned to both organic functional groups hypothesized to have different stabilities (2930 cm-1 (C-H), 1620 cm-1 large (C=O, C=C), 1620 cm-1 (C=C), and 1530 cm-1 (C=O, -COO-)) and mineral associated functional groups (3630 cm-1 (Si-O)) to further elucidate soil properties important for SOC stabilization. Bulk soils were also scanned with visible and near infrared spectroscopy (Vis-NIR) using a portable spectrometer in order to develop multivariate models for prediction of SOC or soil properties important for SOC stabilization.
Relationships between soil properties hypothesized to influence SOM stability and SOC stocks of bulk soil and each fraction were investigated using a multilinear mixed effects modeling approach. To that end, four model structures were developed in which C stocks were modeled using either basic characterization data, mineralogy data, MIRS data, or the most significant variables identified though the three other models as fixed effects and drainage class, physiographic region, or a combination of both as random effects. Additionally, partial least squares regression (PLSR) and support vector machines (SVM) coupled with Vis-NIR spectroscopy were compared to determine the optimal approach to rapidly estimate soil properties important for SOM stabilization. Models were developed to predict extractable Ca2+, Mg2+, K+, extractable acidity, cation exchange capacity (CEC), base saturation, total sand, silt, and clay, fine clay, SOC and permanganate oxidizable C (POXC).
The first study in this thesis showed that, while hypothesized to be important for SOC stabilization, total clay and fine clay were not found to be significantly related either to bulk SOC or any of the separated fractions in Ohio soils. Instead, the results from this study indicate the type of clay is more important than the overall amount, as the relative abundance of illite in the clay-sized fraction was correlated with the stable fractions of si + c and rSOC. Additionally, peak area 3630 cm-1, which is associated with functional groups associated with 2:1 phyllosilicates, was also a significant model factor for the si + c and rSOC fractions. Extractable K+ was significantly and positively correlated with C stocks in bulk, tPOM, and the si + c fractions.
Vis-NIRS predictions presented in the second study of this thesis were assessed using a performance threshold proposed by Chang et al. (2001) in which model performance is assessed by residual prediction deviation of validation (RPDv) (RPD > 2 = excellent, RPD > 1.4 = fair, and RPD < 1.4 = not reliable ). Based on these categories, excellent models were found for Ca2+ (SVM), base saturation (PLSR), base saturation (SVM), CEC (SVM), clay (SVM). fine clay (SVM), SOC (SVM), and POXC (SVM). Fair models were found for Ca2+ (PLSR), Mg2+ (PLSR), Mg2+ (SVM), K+ (SVM), extractable acidity (PLSR), extractable acidity (SVM), CEC (PLSR), sand (SVM), clay (PLSR), fine clay (PLSR), SOC (PLSR), and POXC (PLSR). Not reliable models were found for K+ (PLSR), sand (PLSR), and silt (PLSR). All soil properties were estimated with an R2 > 0.29 and RPD > 1.19. The best performing model was the SVM Ca2+ model (RMSEv = 2.14 cmolc kg-1, RPDv = 2.62, and R2v = 0.85). The poorest performing model was developed with PLSR to predict K+ (RMSEv = 0.11 cmolc kg-1, RPDv = 1.19, R2v = 0.29). Except in the case of base saturation, all soil property predictions improved with SVM models as compared PLSR models.