Remote detection and measurement of leaf-cutting ant nests using deep learning and an unmanned aerial vehicle

Abstract

Leaf-cutting ants are the main group of insect pests in Brazilian forest plantations, and their nests can be visually identified in remotely sensed images. This study compares two distinct pattern recognition methodologies, each with different computational costs, for detecting and measuring leaf-cutting ant (LCA) nests in RGB images acquired by unmanned aerial vehicles (UAVs), aiming to develop an efficient detection system for day-to-day Eucalyptus plantation management. The first methodology (MLP) is based on traditional Multilayer Perceptron Neural Networks combined with the sliding window technique. The second methodology uses the new Convolutional Neural Networks YOLOv5 architecture, which is a deep learning approach applied to RGB imagery and requires more time and memory. For each input image, existing LCA nests were detected and measured based on their bounding box areas. Images were classified to detect the presence and measure the area of LCA nests. The quality of both methodologies were evaluated using accuracy, Kappa, sensitivity, specificity, and mean absolute percentage error (MAPE) metrics, considering training and test data sets. Performances of the YOLOv5 methodology were highly superior to those of the MLP methodology, with 98.45% accuracy using the YOLOv5 large architecture net and 0.49% MAPE in measuring the nests with YOLOv5s nets, whose performance was superior to the MLP results (65.45%), demonstrating high complexity in identifying the targets in the field. The YOLOv5 approach, applied to RGB images acquired by UAVs, shows great promise for precision monitoring of LCA nests, and thus can reduce and optimize insecticide use in forest plantation areas.