Application of machine learning strategies in the detection of invasive plants in pastures

Abstract

The presence of invasive plants on pastures is a challenge for livestock farmers. Weeds not only compete with forage for nutrients, but can also harm livestock. Chemical control is one of the options to tackle this problem. In this sense, automatic detection is a necessary alternative, as it increases agility and reduces process costs. In this sense, machine learning algorithms emerge as possible alternatives. In this work, different models for the classification of invasive plants were evaluated, using images of the genus Rumex as reference. The evaluated models were based on monolithic strategies and on systems with multiple classifiers. The experiments, performed on two datasets with different proportions of positive and negative classes, showed that KNN was the most competent single model in detecting the invasive plant, achieving an accuracy of more than 95% and a sensitivity of more than 0.90 in both datasets. Among the models based on ensembles, Random Forest stood out with a accuracy of over 0.95 and a sensitivity of about 90% in both scenarios.