Nutrient losses from agricultural fields are the largest sources of phosphorus (P) entering the Great Lakes in North America. Research has suggested that multiple conservation practices (CPs) used together (stacked) are an effective way to reduce the amount of P losses individual fields; however, some P loss still occurs. Advancements in the chemical removal of P have provided landowners with an opportunity to capture P that has leaves fields in runoff before it enters local waterways and act as a final polishing agent. A commercially available phosphorus sorbing material (PSM) in the form of a geotextile filter was installed on two well managed fields in midwestern Ontario (ILD and LON) to determine its efficacy in removing dissolved reactive P (DRP), total P (TP) and total suspended solids (TSS) from surface runoff, thereby reducing edge of field P losses. Laboratory tests on unused and used filter material were also conducted to try to determine the sorption potential, amount of P stored in the filter and the mechanisms of P removal. During the two-year study period, the filter removed 0.018 kg/ha of DRP, 0.4 kg/ha of TP and 8.75 kg/ha of TSS at the ILD site. In contrast, the filter at LON released 0.22 kg/ha of DRP and 0.15 kg/ha TP, but removed 37 kg/ha of TSS. The filter most effectively removed P within 8 months of filter installation, suggesting that time was a critical factor impacting performance, among others. Laboratory tests on unused (new) and used (field) filter material indicated that the raw filter material had a large potential to adsorb DRP under controlled conditions, and that this potential was smaller in used material from LON but not ILD. The extraction of P from used filter material indicated that the filters retained approximately 200 mg/kg of DRP at each site (0.0027 kg/ha at ILD and 0.0022 kg/ha at LON) with the majority of this P held in more soluble form likely associated with the metal oxides/clay filter components. This suggests that previously retained P has the potential to be rereleased from the filter. Additionally, the amounts of P held in the filter material calculated via lab tests was considerably lower than the amount removed through the water samples calculations suggesting that some of the P removed by the filter did not stay inside the material. The results of this study demonstrate that P has the potential to be chemically removed at the edge-of-field, but the efficacy of filters as a CP differs in both space and time. This thesis has shown how an edge-of-field filter for surface runoff can be implemented in a field setting in midwestern Ontario, and has identified and which factors are most important to determining the efficacy of this practice.

Soil erosion remains a primary challenge in the 21st century threatening fresh water and cropland that supports more than 95% of global food production. It is of significance to plan for and prevent soil erosion in its initial stages rather than labor intensive repairing later. The Middle Thames River watershed has suffered from severe erosion issues for more than ten years with 21% highly erodible lands throughout the basin, where extensive soil conservation measures are highly encouraged. A series of practical measures that landowners can apply to enhance soil health and water quality while preserving or increasing agricultural production are termed farmland Best Management Practices (BMPs). Among these measures, grassed waterways, as broad and shallow channels to move concentrated surface runoff, are considered as one of the most effective measures to prevent ephemeral soil erosion. Therefore, identifying the site-specific opportunities for grassed waterways implementation in the Middle Thames River watershed can support targeted soil conservation and the watershed planning. This study aims to identify the potential locations for grassed waterways implementation in the Middle Thames River Watershed using four different techniques with high-resolution data (Compound Topographic Index model, Stream Power Index threshold model, weighted linear overlay, fuzzy logic analysis). The Compound Topographic Index model and Stream Power Index threshold model have been developed to predict the existing and potential grassed waterways at the field level. Then the Compound Topographic Index and Stream Power Index threshold models, the multi-criteria decision analysis (MCDA) has been conducted to map the priority areas for grassed waterways implementation at the watershed scale. The output maps of the Compound Topographic Index model and Stream Power Index threshold model display the location and length of predicted grassed waterways in each field. To better visualize the results of the Compound Topographic Index model and Stream Power Index threshold model, the density distribution maps of predicted grassed waterways throughout the studied watershed have been created based on the outputs from Compound Topographic Index and Stream Power Index threshold model. The performance of the Compound Topographic Index and Stream Power Index threshold model have been assessed by visual evaluation, occurrence evaluation and length evaluation. After developing Compound Topographic Index and Stream Power Index threshold models, the multi-criteria decision analysis (MCDA) has been conducted to map the priority areas for grassed waterways implementation at the watershed scale. Twelve factors were selected as criteria of MCDA based on literature review, data availability and geographic knowledge. Two methods including weighted linear combination and fuzzy logic analysis were employed in MCDA, which produced two outputs maps of priority areas for grassed waterways implementation. The results of these two maps have been validated using existing grassed waterways. The results of the Compound Topographic Index model and Stream Power Index threshold model display the existing and predicted grassed waterways in each field. The Compound Topographic Index model with the threshold of 600 has identified 30 existing grassed waterways, while the Stream Power Index threshold model with the threshold of 0.01 standard deviation identified 23 grassed waterways. Several discontinuities exist in predicted grassed waterways along the trajectories of digitized grassed waterways. The lengths of predicted grassed waterways by Compound Topographic Index model have a much better agreement with observation than that of Stream Power Index threshold model. The density distribution map of Compound Topographic Index and Compound Topographic Index model presented high-density areas of predicted grassed waterways which are mainly situated in the northern and central part of the study area, especially the areas along the upstream of Middle Thames River and Nissouri creek. The low-density areas for grassed waterways implementation are mostly located in the southwestern part of the study area. The results of weighted linear combination and fuzzy logic analysis displayed the high-priority areas mainly located in the northwestern part of the watershed, especially along the upstream of Nissouri creek. It is found that these upstream areas have relatively steeper slope gradient than other areas in the studied watershed, with dominant soil type of sandy loam and silty loam. There are more areas belonging to the lowest priority zone and lower areas falling into the most priority level in the fuzzy logic analysis output map, compared with the map of weighted linear combination. The fuzzy logic analysis required less prior knowledge of the relationship among criteria, which provide more flexibility and convenience to decision makers. The validation of both weighted linear combination and fuzzy logic analysis output maps displays relatively good performance, based on the criteria that a greater percentage of grassed waterways implementation must occur in the higher priority zones (Kanungo et al., 2009).

In recent times, with the increasing availability of large datasets, applications of machine learning techniques have grown at a rapid speed. However, due to the black-box nature of these tools, it can be hard for model builders to understand the detailed structure of the system that machine learning models simulate. Agent-based modelling (ABM) is a popular approach to studying complex systems., One of the challenges for this technique is to design the decision making processes of the agents in the model. As machine learning tools have a strong ability to transform the information from the raw data into a functional model as the decision making processes for agents in ABMs. Because an ABM can provide a detailed structure for the system that the machine learning model simulates, it is reasonable to combine the two kinds of models. However, although in previous studies, some researchers combine the two models, most of them use one of the two models as a validation tool for the other, rather than to integrate the machine learning model into the decision making processes of agents in ABMs. Therefore, this thesis focuses on integrating a machine learning model into the ABM, and contrast it with the ABMs with two traditional decision making models, including an optimal model and a stochastic model. To compare the three decision making models, we use farmers’ BMP adoption case in the Upper Medway subwatershed, and contrast the three models through three metrics, including the percentage of BMP adoption, size of agricultural land of BMP adoption, and the correlation between BMP adoption and landuse types. As a result, the ABM with the machine learning model presents a high level of accuracy compared with the other two traditional models, but its adaptability to other cases and the robustness to uncertainties still require a further study.