Supervision:

David Makowski (INRAE) and Fernando Miguez (Iowa State University).

Thesis presented and defended on 09 December 2020

Summary:

An on-farm research network is an organization of farmers that conducts agronomic experiments under local conditions. There is growing interest in on-farm research networks because they provide the infrastructure needed to test new products and management practices in farmers’ fields. Often, the results are usually presented as individual reports (i.e., a report summarizing the outcome for one trial), but this provides limited information difficult to generalize and does not allow presenting, in a synthetic way, all the results collected from the different trials. Moreover, there is an unexplored potential in detecting yield response variability patterns for better decision making. The overall objective of this thesis is to demonstrate the importance of identifying appropriate statistical methods for analyzing and visualizing on-farm research network data. Specifically, I focused on analyzing the on-farm research networks managed by the Iowa Soybean Association, and an adaptation was made with a French case-study. A data-analytics framework was developed to analyze multiple trials that use a common protocol and identify the conditions where an imposed treatment may or may not be effective. This framework used a random-effect model through a Bayesian approach and returned yield response estimates at the network and trial levels. The framework was implemented through a web-application for 51 different management practices on corn and soybean. The web-application includes dynamic data visualization features to enhance communication and information sharing, and is accessible to a broad audience to improve accessibility to on-farm research insights. A random-effects statistical model was used to compute prediction intervals describing a range of plausible yield response for a new (out-of-sample) trial, and compute the probability that the tested management practice will be ineffective in a new field. Depending on the level of between-trial variability, the prediction intervals were 2.2–12.1 times larger than confidence intervals for the estimated mean yield responses (i.e., at the network level) for all tested management practices. Using prediction intervals and the probability of ineffective treatment will prevent farmers from over-optimistic expectations that a significant effect at the network level will lead with high certainty to a yield gain on their farms. The data-analytic framework was adapted to a French on-farm research network focusing on the efficacy of biocontrol agent products against Botrytis cinerea, potassium bicarbonate and Aureobasidium pullulans, on organic vine. The results favored potassium bicarbonate as its efficacy on incidence at the network level is higher for diseased intensities between 0% and 10% than for Aureobasidium pullulans. For both biocontrol agents, the efficacy on incidence for a new trial is highly uncertain for intensity levels higher than 15%. Finally, this research investigated the impact of experimental plot scale (i.e., small-plot scale and field scale) on the effect of management practice on crop yield and identified the cause of potential discrepancies to inform on-farm decision-making better and adapt the extrapolation of the results. Taken together, this research represents the first major effort in consolidating results from on-farm research network and provides insight to make better farming management decisions.

Anabelle Laurent CV