Soils are a non-renewable resource, which is the foundation of all ecosystems. Mismanagement of soil particularly in agro-ecosystems has degraded soil. To guide management of soils, to remediate soils, and enable optimal agricultural production, soil health indicators are needed.
The objective of this dissertation was to determine the potential of biological and other soil properties to predict soybean yields. The central approach was based on soil samples from farmers’ fields instead of long-term experimental sites (LTES). Farmers’ fields in this study represented diverse management practices that exist in the agricultural sector. Soil Health (SH) measurements that are calibrated and that can consistently detect land management are lacking which was shown in Roper’s et al. (2017) 2017 publication that found existing SH tests (CASH, Haney) had limited ability to identify agronomic land management practices at a North Caroline LTES. And that they were poorly correlated with crop yields. This means that the quote by the Soil Health Institute “There is no standardized measurement for Soil Health in the United States” is still true.
Extensive research has found certain soil enzyme assays to be quite sensitive for detecting land management effects and exhibit seasonal stability. The currently promoted SH indicator scores have limited or inappropriate biological indicators (e.g. microbial biomass and respiration). The latter measurements vary too much on a seasonal basis due to weather variation or a recent short term soil management event (e.g. high organic inputs, disturbance). Thus, the global objective of this dissertation was to determine the potential of biological soil properties, specifically enzyme activities and microbial community biomarkers to quantify SH. Enzyme activity has the added advantage over most other soil biological measurements, that it can be run on air-dried soil and furthermore is relatively simple. This is attractive to commercial labs who want to minimize analytical costs and prefer to use air-died soils. A secondary objective was to determine the potential of soil properties to predict soybean yield with an individual variable or in a multivariate model.
Unlike most previous research on SH, which was based on data from long-term experimental sites (LTES), this investigation utilized analyses of soil samples from farmers’ fields. Farmers were surveyed to collect historical management information on each field after which LIDAR data, and soil type information from the Soil Survey website was obtained. For the 2019, 2020, 2021 growing seasons Ohio sampling sites were visited during the spring season and a composite soil sample at a depth of 0-15 cm was collected. In 2021 soil samples were collected at three LTES in Ohio and one in Michigan. Furthermore, soils were sampled at two virgin and two restored prairie sites in Ohio.
In fall of each year, on farm fields that grew soybeans, soybean yields were determined at each soil sampling site. Soil samples were analyzed for microbial community composition, enzyme activities, total carbon (TC), soil organic carbon (SOC), total nitrogen (TN), pH and texture. Microbial communities were profiled using the Ester-Linked Fatty Acid Methyl Ester (EL-FAME) analysis. The enzyme activity of β-glucosidase (NAG), N-acetyl glutamate synthase (NAG), and arylsulfatase (AS) were chosen because previous research has shown these measurements have been shown to be sensitive in detecting soil/crop management effects.
Chapter 1 of this dissertation reviewed the literature on SH.
The objective of Chapter 2 was to determine if enzyme activities, individual soil properties, or other variables could predict soybean yield. The study used soils from farm and long-term experimental sites (LTES) in Ohio and one LTES in Michigan. The first statistical analysis was used to correlate these variables with soybean yields. The data was further evaluated separately for conventional and organic land management. Previous research has shown organic management has lower yields. This is because organic production requires wider rows to enable mechanical weed control, as herbicides are not allowed for certified organic production.
The global objective of Chapter 3 was to develop a biochemical Soil Health Index (SHI). To develop this index, multivariate soybean yield prediction models were rated for their fitness based on R2 values generated with a general and generalized linear analyses. To evaluate the strength of specific variables, a stepwise decrease of input variables was conducted in the model development. Instead of using individual variables from the multivariate model, data was put into 7 categories (land management, soil texture, environmental factors, total nitrogen, soil org. carbon, enzymes, and EL-FAME). All 7 categories represent 105 variables. Each variable was converted to a relative value constrained from 0 to 1 based on the maximum value for a given variable. To reduce the mean squared error produced by generalized linear model (GLM), a least absolute shrinkage and selection operator (Lasso) regularization step in combination with a cross-validation was performed. To automate this process the package glmnet was used in RStudio. This statistical analysis resulted in a multivariate model that accounted for intercorrelations between variables and that was more robust because it was cross validated by running more than 10000 possible combinations of training vs. test data sets.
The resulting soybean yield prediction model had a R2 value of 0.84. To develop the biochemical SHI the individual slope coefficients were separated into negative and positive groups. This information was used to calculate the weighted variable value by taking the total slope coefficient and multiplying the individual assigned weight factors with each variable data point. Because the variables and the weight factors have a range of 0 to 1, each observation resulted in a SHI score with the same range (0 to 1) after the weighted variable scores were summarized. These calculation steps were conducted separately for the positive and negative slope coefficient which included EL-FAME and enzyme coefficients. Additionally, SHI scores were determined separately for EL-FAME variables and enzyme variables which is different to the original SHI that used EL-FAME and enzyme variables.
The most common SH indicators in this study and the computed SH scores were analyzed for their ability to detect soil management at four LTES by running the Tukey's Honest Significant Difference test in combination with a sensitivity scoring algorithm. The sensitivity scores were used to identify the most sensitive SH indicators.
In Chapter 4 a total of 521 soil variables were scored for their ability to detect agricultural land management practices (e.g. crop rotation, cover cropping, soil amendments, tillage practices), restored prairies, and virgin soil in Ohio. Additionally at the agricultural scale each variable was tested for its temporal sensitivity. The most sensitive SH indicators were identified with the help of a sensitivity scoring algorithm and their relationship to soil organic carbon was determined. The remaining SH indicators were used to determine beneficial and detrimental agricultural land management practices.