Development of Simple Bayesian Belief and Decision Networks as Interactive Visualization Tools for Determining Optimal In-row Spacing for ‘Beauregard’ Sweetpotato

Abstract

A Bayesian belief network (BBN) model, which we named BxINROW/NET, was developed to represent the relationships among in-row spacing regimes, some agroclimatic variables known to influence storage root initiation, growing degree-days (GDDs) to harvest, and yield grades in ‘Beauregard’ sweetpotato grown in Louisiana. The model was developed from experimental data collected in a subset of years between 1990 and 2010 and assumed that soil moisture, weeds, and chemical injury were not limiting variables during the growing season. The BBN model error rates for storage root yields were 21%, 20%, and 13% for U.S. #1 (US1), canner, and jumbo grades, respectively, as estimated from repeated random partitioning of the modeling data set into training and testing partitions. In comparison, the error rates for a baseline logistic regression model were 56%, 54%, and 53% for US1, canner, and jumbo grades, respectively. The BBN model showed that GDDs to harvest (GDDH) as well as air and soil temperatures during the critical storage root initiation period [20 days after transplanting (DAT)] interacted with in-row spacing regimes to help determine the yield outcomes. Under a uniform irrigation management and minimum to intermediate GDDH (980 to 1495 GDDs), narrow (20 to 22 cm) to intermediate in-row spacing regimes (30 cm) were associated with higher probabilities (56% to 71% of cases) for attaining a high US1 yield (22 to 45 t·ha−1). These outcomes were associated with minimum to intermediate soil and air heat units 20 DAT, representing early to intermediate planting dates. Under similar conditions, wide in-row spacing treatments (38 to 40 cm) were associated with increased probabilities (100% of cases) for achieving a high yield of jumbo or oversized roots if GDDH (1495 to 1710 GDDs) was maximized. BxINROW/NET was also used as the foundation model to construct Bayesian decision networks (BDNs) for fresh market and processing scenarios. The BDNs were constructed by adding a value or gain node associated with each yield grade. Nodes representing price per box and stand deficiency were also added. These nodes allowed the prediction of estimated net return associated with a specific in-row regime given some agroclimatic variables and GDDH. As a result of its reliance on conditions observed in the study, BxINROW/NET is only applicable to a local Louisiana growing area. Further study is necessary to determine the model's applicability in other regions and growing conditions.