Automated lepidopteran pest developmental stages classification via transfer learning framework.

Area of Science:

• Agricultural Science
• Computer Science
• Entomology

Background:

• Maize is vulnerable to lepidopteran pests, complicating manual identification and control of larval stages.
• Accurate identification of pest larval instars is crucial for effective integrated pest management (IPM).

Purpose of the Study:

• To develop and evaluate an automated classification system for identifying larval developmental stages of four key lepidopteran pests of maize.
• To compare the performance of five Convolutional Neural Network (CNN) architectures for this classification task.

Main Methods:

• Utilized five CNN architectures (ConvNeXt, Densenet121, EfficientNetv2, MobileNet, ResNet) fine-tuned with two optimizers (SGD with momentum, Adam).
• Evaluated models based on accuracy, precision, recall, and F1 scores for classifying 23 instars across four pest species.
• Tested the best performing model in a natural field environment.

Main Results:

• The Densenet121 model with the Adam optimizer achieved the highest classification accuracy of 96.65% in laboratory settings.
• This model demonstrated high performance metrics: 98.71% precision, 98.66% recall, and 98.66% F1 score.
• In field tests, the Adam_Densenet121 model achieved 90% accuracy in identifying the larval stages of the four pests.

Conclusions:

• Transfer learning-based CNN models, particularly Densenet121 with Adam, effectively automate the identification of lepidopteran larval instars in maize.
• This automated system offers a promising tool for enhancing precision-integrated pest management strategies in agriculture.