Crop pest image classification based on improved densely connected convolutional network.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Crop pests significantly impact agricultural yield and quality.
• Accurate pest identification is crucial for effective crop management strategies.
• Existing datasets and models have limitations in pest classification accuracy.

Purpose of the Study:

• To develop a robust pest identification model to address data limitations and improve classification accuracy.
• To introduce the HQIP102 dataset for large-scale crop pest research.
• To enhance the DenseNet architecture for superior pest recognition.

Main Methods:

• A curated dataset, HQIP102, was created with 47,393 images across 102 pest classes.
• The MADN model was proposed, integrating Selective Kernel units, Representative Batch Normalization, and ACON activation into DenseNet.
• Ensemble learning was employed to construct the final MADN model.

Main Results:

• MADN achieved 75.28% accuracy and 65.46% F1Score on the HQIP102 dataset.
• Performance improvements of 5.17% (accuracy) and 5.20% (F1Score) over pre-improved DenseNet-121 were observed.
• MADN outperformed ResNet-101 with 10.48% higher accuracy and 10.56% higher F1Score, while reducing parameter size by 35.37%.

Conclusions:

• The MADN model demonstrates superior performance in crop pest identification compared to existing models.
• The HQIP102 dataset provides a valuable resource for advancing pest identification research.
• Cloud-deployed pest identification models can significantly contribute to crop yield and quality preservation.