InsectNet, supported by a vast collection of 12 million insect images—many gathered by citizen scientists—offers identification and predictions for over 2,500 insect species with an accuracy exceeding 96%.
A farmer spots an unusual insect on a leaf.
Is it a pollinator? A pest? Will it bring good news at harvest? Or is it something that needs control?
The farmer can take a photo and use a smartphone or computer to upload it to a web application named InsectNet, which utilizes machine learning to provide immediate insights.
“The app not only identifies the insect but also predicts its taxonomic classification and its role in the ecosystem, whether it’s a pest, predator, pollinator, parasitoid, decomposer, herbivore, indicator, or an invasive species,” noted a recent scientific article about InsectNet published in the journal PNAS Nexus. The study is led by Iowa State University’s Baskar Ganapathysubramanian and Arti Singh.
With its backing of 12 million insect images from various sources, including citizen scientists, InsectNet can accurately identify and predict information for more than 2,500 insect types. If the app is uncertain about an identification, it communicates that to the user, thus enhancing their confidence in the provided answers.
Furthermore, designed to adapt from global to local contexts, the app can be refined using region-specific, expert-verified datasets, making it beneficial for farmers across different areas.
So, watch out for harmful insects like armyworms and stink bugs, as well as welcome the friendly pollinators like butterflies and bees, and the pest predators like lady beetles and mantises.
“Our vision for InsectNet is to enhance current methods and be part of an expanding array of AI tools designed to tackle agricultural problems,” the authors stated.
A community of researchers
The capability to customize InsectNet for specific regions makes it especially valuable, according to Singh, an associate agronomy professor.
In Iowa, for example, Singh pointed out that about 50 insect species are particularly significant for the agricultural sector. To help identify and provide predictions about these insects, the project utilized around 500,000 insect images.
This could be replicated for farmers worldwide. And even in regions where data is scarce—often requiring millions of images for local fine-tuning—there is still access to a global dataset for farmers.
However, InsectNet isn’t exclusively for farmers; Singh mentioned it could also aid port and border agents in identifying invasive species, as well as assist researchers in ecological studies.
The app is functional and adaptable, but is it accessible?
Currently, you can’t download it from an app store, noted Ganapathysubramanian, the Joseph and Elizabeth Anderlik Professor of Engineering and director of the AI Institute for Resilient Agriculture at Iowa State. Nevertheless, the app runs on a server at the university. By using a QR code or this link (insectapp.las.iastate.edu/), users can upload insect images for identification and predictions.
This application functions effectively throughout all life stages of an insect—from egg to larva to pupa to adult. It handles similar-looking species and adapts to various image quality and angles.
The essential query for users is straightforward: “Is this a pest?” Singh emphasized. “Or is it beneficial?”
Developers showcased the app during the Farm Progress Show in Boone, Iowa, last August, and now the research article is aimed at introducing it to a wider scientific audience.
But aren’t there existing apps for insect identification?
Indeed, Ganapathysubramanian confirmed, but they don’t match the scope of InsectNet and lack capabilities for global-to-local applications. Plus, they aren’t open-source, meaning their technology isn’t readily shareable.
“By making InsectNet open source, we can stimulate wider scientific engagement,” he stated. “This allows the scientific community to build upon these initiatives instead of starting from the beginning.”
The project addressed numerous technical inquiries applicable to other initiatives, he added.
What constitutes sufficient data? How can we acquire that quantity? How should we handle inconsistent data?
What level of computational power is required? How do we manage large data volumes?
“Ultimately, it takes a team effort to reach this stage, right?” Ganapathysubramanian enthused.
It took a collaborative effort of agronomists, computer engineers, statisticians, data scientists, and AI experts approximately two years to develop and optimize InsectNet.
“What we gleaned from working with insects can extend to include weeds, plant diseases, or any related identification and classification challenges in agriculture,” Singh mentioned. “We’re nearing a comprehensive solution for identifying all these factors.”
