Unconventional strategies for aphid management using smart computer vision models in sorghum

Abstract

Aphids are a challenging crop pest to manage. The sorghum aphid, for example, causes significant yield loss in unmanaged sorghum. One of the key strategies to manage this pest and prevent yield loss is using pest monitoring to determine an economic threshold level used for spraying insecticides. However, pest monitoring is time-consuming and requires trained personnel and regular visual assessments across large field areas once aphids are detected on sorghum plants. There is a need to develop alternative methods to monitor and manage the diverse various pests on crops effectively. The economic impact of the sorghum aphid, Melanaphis sorghi (Theobald) (Hemiptera: Aphididae), their small size and high reproductive rate make it a suitable study organism to examine new technology. The technology should automate the visual assessments performed during pest monitoring, and the further development of this technology using new management strategies compared to traditional management on field crops. Therefore, the objective was to understand the use of technology to develop automate aphid visual assessments and develop new management strategies, which can be applied with the further development of this technology. A series of computer experiments and field trials were conducted to understand how technology, specifically machine learning, can be applied to aphid identification and the responses of different management strategies based on economic threshold levels. Computer experiments demonstrated that machine learning using deep learning can classify different aphid densities on images found on standard economic thresholds for spraying with an accuracy of 86% and the correct classification of aphids’ densities as above or below threshold spray density over 97% of the time. Together this information suggested that deep learning can be applied to the automated classification of aphid densities found on images and further development of detection and counting tasks of aphids, usually common activities performed during pest monitoring of aphids. Consequently, the examination of detection and counting aphid individuals on images based on standard densities for spraying was performed using the same technology (i.e., deep learning) with different image processing techniques. Results showed 92% precision, and 21% mean percent error of miscounting aphids on images. This helps to inform that automation can be applied to pest monitoring of aphids, potentially reducing the time for visual assessments of pests and training personnel. The combination of automated classification-detection of aphids can lead to the identification of other insect organisms commonly found in sorghum environments. These insects, primarily coccinellids, are considered beneficial insects that can manage aphid populations naturally. Therefore, we used similar approaches to detect and classify common coccinellids found in sorghum. Together this work demonstrates that deep learning can be implemented in pest monitoring of aphids creating a smart insect monitoring system that can lead to new strategies for making management decisions. The new suggested strategies was tested on field trials to understand the responses of sorghum aphids and coccinellids under commonly used spraying strategies for aphid management, including economic and tally threshold levels and new strategies on randomly and specific spraying plants on whether or not aphids are present on plants. Field experimental results showed that we can manage aphid populations by applying insecticides to individual plants with or without aphids, reducing the environmental effects and pest resistance. Collectively this information will aid in the use and adoption of deep learning models for pest management and the development of mobile applications and unmanned vehicles with sophisticated sensor systems to manage aphids with the new proposed pest strategies.