Machine learning and microscopy could change the future of insect identification
Accurate insect identification is the first and most important step in pest control. Insect identification is essential in agriculture to choose the best management approaches.
Tobacco bollworm (Helicoverpa zea) and budworm (Chloridea virescens) both hatch five days after oviposition. Helicoverpa zea has developed tolerance to Bt transgenic crops, therefore farmers must apply insecticides during the egg-laying window. Helicoverpa zea eggs are tiny, measuring 0.5 mm in diameter, making it necessary to use a microscope to distinguish the different species.
A recent study published in Agriculture suggests a machine learning approach using a convolutional neural network to classify different types of caterpillars with high accuracy. The researchers created a dataset of approximately 5,500 images to quickly train and test the network using the Multi-Camera Array Microscope (MCAMTM), a gigapixel-sized parallelized microscope and an automated image processing pipeline. In the future, the software could allow farmers to take an image of eggs on a leaf and instantly identify the species before the eggs hatch.
Damage caused by uncontrolled insects
The tobacco budworm Chloridea virescens (formerly Heliothis virescens) and the bollworm Helicoverpa zea (Lepidoptera: Noctuidae) are important agricultural pests. When left unchecked, they cause most of the pest damage to cotton in the southeastern United States.
The larvae consume various agricultural products, such as cotton, corn, tobacco, soybeans, tomatoes, wheat, and vegetable garden plants. They are managed in cotton using transgenic plants that express the insecticidal Cry proteins of the bacterium Bacillus thuringiensis (Bt), together with the occasional application of chemical foliar insecticides. Unfortunately, Cry protein resistance is spreading and getting worse.
Importance of insect egg identification in pest control
Identification is the first step in managing the US tobacco bollworm-budworm complex in cotton. This allows for the adoption of the appropriate management approach for that particular insect.
The economic threshold for bollworms in cotton was based on the presence of live larvae or damaged plant reproductive tissue prior to Bt resistance in Helicoverpa zea. After Helicoverpa zea developed resistance to Bt, egg-based thresholds were adopted to decide when to spray.
Challenges in Identifying Insect Eggs
Heliothis virescens and Helicoverpa zea are identical in color, size and shape. The only difference between them is the presence of small cuticular ridges on the surface of Helicoverpa zea. Egg identification requires a high magnification brightfield microscope and a researcher with extensive experience in determining the small cuticular ridges on the surface of the egg.
When managing populations of common pests, this approach is not practical because it is necessary to collect many samples from fields spread over a large geographical area.
Development of a dataset of Heliothis Virescens and Helicoverpa Zea using machine learning methodology
Deep learning is used more in the biological sciences. Applying convolutional neural networks (CNN) to digital photographs produces repetitive image classifications. An image dataset of Heliothis virescens and Helicoverpa zea was developed by Efromson et al.
The researchers used a multi-camera microscope (MCAMTM) and, using the data from this analysis, created a machine learning methodology to quickly distinguish species.
The subtle and difficult physical changes between two species make this the perfect test case to demonstrate the usefulness and effectiveness of deep learning in identifying insect eggs. Rapid, unbiased, high throughput, reproducible analysis, low cost, and a means to improve the sustainability of insect management are all advantages of this technique.
The future of pest control using deep learning
This research aims to demonstrate the viability of egg identification through high-resolution microscopy and machine learning. A computer vision methodology would offer readily available, reproducible identification that otherwise would not be available for the short time needed to make pest control decisions. For example, one option might be to take a high-resolution photo of an insect egg with a smartphone to identify the type of insect.
The researchers found that a machine learning method had a 99.3% accuracy rate in classifying egg images of two caterpillar species. The two species studied represented the worst case situation in which only minute morphological variations exist between species. In the past, identifying these caterpillar eggs required the use of species-specific antibody assays or microscopy to examine minute morphological variations.
The suggested deep learning algorithm only required a representative high-resolution digital image of the user. It generated automated and unbiased output, unlike the labour-intensive subjective analysis required by current methods. This is the first account of an insect species being identified from eggs using deep learning techniques, which has implications for other important pest management issues.
Efromson, J., Lawrie, R., Doman, TJJ, Bertone, M., Bègue, A., Harfouche, M., Reisig, D. and Roe, RM (2022) Identification of caterpillar egg species by l machine learning using a convolutional neural network and a massively parallelized microscope. Agriculture, 12(9), 1440. https://www.mdpi.com/2077-0472/12/9/1440/htm