How artificial intelligence could make it easier to identify common corn pests

By Andrew Porterfield

Corn is an economically important crop in Asia, Europe and America. But it’s not just people and farm animals who consume corn. The crop is also eaten by more than 30 species of insects of the order Lepidoptera (i.e. moths and butterflies). Four major pests are the most destructive: Ostrinia furnicalis, Spodoptera frugiperda, Mythimna separataAnd Spodoptera litura.

The destructive abilities of these insect pests vary depending on the larval stage or instar. Therefore, it is important for farmers and pest controllers to be able to identify the stages of each pest in order to optimize their control. Traditionally, identification is done with the naked eye and knowledge of each individual pest. This can be slow and tedious and can result in butterfly stages being misidentified. There is very little difference in size between stages and the larvae of different species have similar colors.

A research team from Jilin Agricultural University in China recently investigated whether certain artificial intelligence (AI) models could identify the different stages of these insects more quickly and accurately. After testing several AI models, they found an AI model and data optimizer combination that was most effective for classifying insects by age across 23 developmental stages of the four pests. The researchers found that this combination of AI and machine learning had an overall accuracy of 96.65 percent, outperforming all other models they tested. Their results were published in October Environmental entomology.

By using machine learning algorithms, farmers and researchers can make identification faster and easier and detect pests earlier, enabling more precise control measures and improving both crop protection and sustainability. However, it is important to use the right AI model.

A key problem in applying AI to insect identification is being able to manage images from different sources and different sizes and convert this raw data into a usable and shareable representation of insect characteristics such as shape, geometry, color and texture.

Ri-Zhao Chen, Ph.D., and his team at Jilin Agricultural University used convolutional neural network (CNN) models that can better classify and recognize images while requiring less memory and computational effort. They also applied transfer learning, which is based on pre-trained models of previous machine learning tasks, further reducing the need for advanced (and expensive) computations.

In their study, the researchers selected larvae of the four most damaging lepidopteran corn pests (O. furnicalis, S. frugiperda, M. separata, And S. litura). The larvae’s only source of food was corn leaves. Twenty larvae were selected for rearing. The first to fourth larval instars of all four species were photographed with a microscope, while the fifth larval instar of O. furnicalis and the fifth and sixth instars of the other three species were photographed with a mobile camera. The researchers took 20 images of each pest every 24 hours. During the field tests, the researchers placed insect larvae in an existing corn field and took pictures there with a cell phone.

So the main task of the AI ​​was to break down all this different data and correctly identify each species and stage. Out of five CNN models, one called “Densenet121” stood out as the most accurate. However, the model relied on data expansion software that could harmonize the photos and micrographs. An optimization algorithm called “Adam” proved to be the most accurate.

Artificial intelligence in agriculture is not a new phenomenon, but researchers say the real challenge lies in finding the best model among a “variety of AI-driven approaches.”

“AI-powered systems have enabled the development of precise sprayers capable of delivering optimal herbicide doses directly to targeted weeds,” the researchers write. “In addition, innovations such as spray zones provide invaluable insight into field heterogeneity and optimize insecticide application strategies.”

However, “it remains a challenge to accurately identify and classify the developmental stages of these pests.” Our deep learning-based methods close this gap.”

Andrew Porterfield is an author, editor and communications consultant for academic institutions, companies and non-profit organizations in the life sciences sector. He lives in Camarillo, California. Follow him on Twitter at @AMPorterfield or visit his Facebook page.

Related