Multiple model visual feature embedding and selection method for an efficient pest classification supporting...

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Agriculture 5.0 relies heavily on crop cultivation, necessitating efficient methods for pest identification.
• Manual pest identification is labor-intensive and prone to errors, impacting crop yield and quality.
• There is a growing need for automated, resource-efficient systems for remote pest detection.

Purpose of the Study:

• To develop an automated system for insect pest identification with high efficacy and low resource requirements.
• To create a classification model capable of handling a large number of pest classes.
• To improve the speed and precision of pest identification compared to manual methods.

Main Methods:

• Utilized pretrained deep learning models (DenseNet201, EfficientNetB3, InceptionResNetV2) for visual feature extraction.
• Applied Linear Discriminant Analysis (LDA) for efficient feature selection from combined pest datasets (19 classes).
• Deployed a lightweight dense neural network for final classification.

Main Results:

• Achieved 99.99% accuracy, 100% validation, and 99.99% recall with negligible loss.
• The proposed hybrid feature selection method proved more computationally efficient than traditional transfer learning.
• The system demonstrated high efficacy with reduced computational and memory demands.

Conclusions:

• The developed system offers a low-resource, high-efficacy solution for multi-class insect pest classification.
• This approach is well-suited for deployment in precision agriculture environments.
• The hybrid feature selection strategy enhances classification efficiency without extensive retraining.