COMPARISON OF ARTIFICIAL INTELLIGENCE TECHNIQUES IN THE IDENTIFICATION AND COUNTING OF MACAÚBA PLANTS

Macaúba (Acrocomia aculeata) is a palm tree native to Brazil and its fruits can be used in the manufacture of biodiesel, by-products for animal feed and in the cosmetics industry. The main advantage of macaúba over other alternatives is its wide geographical distribution and endemism in Brazil, with a significant concentration of native massifs in Minas Gerais. For the identification and counting of natural massifs, the use of mapping technologies becomes an effective alternative, especially in natural areas. Thus, the objective was to compare Machine Learning (Random Forest, KNN and Decision Trees) and Deep Learning (You Only Look Once) techniques to evaluate the best performance among the methodologies applied in the identification and counting of macaúba plants. The algorithms were used for the classification of five classes (macaúba, vegetation, pasture, exposed soil and shade) in visible spectrum (RGB) images acquired by an unmanned aerial vehicle (UAV). For deep learning, YOLOv4 was used, whose training was carried out with 2000 and 5000 epochs. From 24 images containing a total of 160 plants, the data were divided into 70 and 30% for training and testing, respectively. Machine learning techniques underperformed when compared to deep learning training. KNN obtained the best overall accuracy (70%), but with 30% accuracy for classification. On the other hand, YOLO achieved accuracy of 63% and 77% with 2000 and 5000 epochs, respectively