Analysis and Prediction of the Quality of Biocontrol using Machine Learning Classifications

Weed control relies mainly on integrated control methods of preventive, agricultural and chemical methods. On pasture lands, however, the chemical methods of spraying pesticides in large area is expensive, has negative consequences on ground water, on environment and on health in general. A safer and more cost effective alternative is biocontrol of weeds in which harmful and unwanted grass, weeds in general are subjected to some natural enemy to control it directly and indirectly. Leafy spurge is one common weed native to central and southern Europe that have spread across western Canada and North America. Not only does this invasive alien plant expand to overtake nearby areas; the milky liquid from its stems and flowers causes severe skin rashes or irritation in in livestock and humans. The weed has been targeted by beetles from the flea beetle genera Aphothona as biocontrol since they were introduced into Canada in the 1980s. It has been discovered that the growth of the A. n. agent and its effectiveness as a biocontrol agent is determined by the interaction of a variety of factors. However, understanding the nature of the relationships among those many factors is incomplete and unclear. A machine learning approach to the analysis of such factors and to the prediction of suitability and potential success of control sites is the subject of this paper. The methodology was used to analyse the available data taken from Regina Agriculture Station weed control project to provide scientists the ability to predict the suitability of sites and the potential success of the agent before its release. It can be also used for the evaluation of existing sites. A number of machine learning classifier algorithms have been adopted and applied to the data including Random Frost, Nearest Neighbour, Support Vector Machines (SVMs), Logistic Regression, Neural Nets and Bayes with variable degrees of accuracy. The adopted classifiers are evaluated with best ones are selected based on Matthew?s correlation factor (MCC) and the overall accuracy of prediction