Computer vision model for sorghum aphid detection using deep learning

Abstract

Aphids are a challenging crop pest to manage. The sorghum aphid, for example, causes considerable yield loss in unmanaged sorghum. One of the key strategies to mitigate yield losses caused by this pest includes monitoring productions fields and using economic thresholds to spray insecticides. However, monitoring aphids is a time-consuming task and requires regular, visual assessments across large hectarage once aphids are detected on sorghum plants. To address this challenge, we propose to use object detection models based on deep learning to automatically detect aphid infestations on sorghum leaves using digital images. We used 1190 images collected during field monitoring events and evaluated the performance of 3 deep learning detection models within the YOLOv5 family that vary in complexity: YOLOv5n, YOLOv5s, and YOLOv5m. We then tested three different image sizes, including input resolutions of 416 × 416, 640 × 640, and 1280 × 1280 pixels. We trained models to detect individual aphids, which ranged between 1 and 125 sorghum aphids/leaf and is comparable to threshold levels used to manage aphids in field conditions (i.e., 50–125 aphids per leaf). Detection models had a precision of 92% precision with a 84.5% recall and 90.6% mAP@0.5 for YOLOv5m Pytorch, making it a potential candidate for quantifying aphid densities using deep learning. The models tested and methodology developed here can be implemented in management decisions of sorghum aphids or as sampling tools for use in screening insect-resistant varieties. Development of mobile applications and integration into unmanned vehicles with sophisticated sensor systems will aid in use and adoption of computer vision models for pest management.