GradCAM-PestDetNet: A deep learning-based hybrid model with explainable AI for pest detection and classification

Area of Science:

• Agricultural Science
• Computer Science
• Ecology

Background:

• Pest detection is vital for food security, agricultural productivity, and economic development.
• Traditional pest detection methods are often slow, inaccurate, and require expert knowledge.
• Advancements in AI and computer vision offer potential for more efficient pest detection systems.

Purpose of the Study:

• To develop an efficient and interpretable pest detection system using deep transfer learning models.
• To evaluate the performance of various object detection and transfer learning models for pest identification.
• To enhance model interpretability through Gradient-weighted Class Activation Mapping (Grad-CAM).

Main Methods:

• Utilized object detection models (YOLOv8n, YOLOv8s, YOLOv8m) and transfer learning techniques (VGG16, ResNet50, EfficientNetB0, MobileNetV2, InceptionV3, DenseNet121).
• Explored Vision Transformers (ViT) and Swim Transformers for complex pattern processing.
• Integrated Grad-CAM for visualizing model predictions and improving interpretability.

Main Results:

• The YOLOv8n model provided the fastest inference at 1.86 ms/img, suitable for low-resource environments.
• An ensemble model (ResNet50, DenseNet, MobileNet) achieved 67.07% accuracy, 66.3% F1-score, and 68.1% recall.
• This represents a significant improvement over the baseline CNN's 21.5% accuracy, indicating a more generalized and robust model.

Conclusions:

• GradCAM-PestDetNet offers a viable and interpretable solution for automated pest detection.
• The integration of deep transfer learning and Grad-CAM enhances detection accuracy and model transparency.
• This AI-driven approach supports efficient pest management in agriculture and ecological studies.