REAL TIME PEST DETECTION USING YOLOv5

Abstract

Depending on the increasing population and nutritional needs, we should develop new methods and systems in agricultural production that take environmental issues into account and ensure efficiency and sustainability. Inappropriate pest control methods can result in 70% of yield loss. The caterpillar is a pest that can be invasive and can damage yield by eating the leaves, shoots, fruit and flower parts of plants and trees. Pesticide spraying is the most preferred pest control method due to its speed of action and scalability. However, due to the increasing environmental and health awareness, less pesticide use is required. One of the important methods of reducing pesticide usage is to spray only the places where they are needed. In order to perform spot spraying, first of all, the location of the pest must be determined. It is possible to detect pests using computer vision methods. In the study, we developed an object detection system to detect the thistle caterpillar (Vanessa cardui), which is encountered in Turkey and can cause damage to sunflower cultivation, in real time via video using the YOLOv5 object detection architecture. For this purpose, we used 2416 images taken under different lighting and background conditions. We trained the object detection system in two different ways using transfer learning and learning from scratch methods and compared the results. Results indicate that the system is functional and being able to correctly detect the thistle caterpillar at 65 FPS.