Classification of field wheat varieties based on a lightweight G-PPW-VGG11 model

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Wheat variety classification is crucial for agriculture but challenging due to subtle morphological differences.
• Traditional methods are inefficient, and existing models are too complex for mobile deployment.
• Intelligent agricultural management requires efficient and accurate variety identification systems.

Purpose of the Study:

• To introduce G-PPW-VGG11, a novel lightweight convolutional neural network for wheat variety classification.
• To address the computational complexity and memory-intensive nature of existing models.
• To enable effective deployment on mobile devices for real-world agricultural applications.

Main Methods:

• G-PPW-VGG11 combines partial convolution (PConv) and partially mixed depthwise separable convolution (PMConv) for reduced complexity.
• Incorporates ECANet for efficient channel attention, enhancing leaf feature capture and noise suppression.
• Replaces VGG11 fully connected layers with pointwise convolutional layers and global average pooling for efficiency.

Main Results:

• G-PPW-VGG11 achieved 93.52% classification accuracy with only 1.79MB memory usage.
• Demonstrated a 5.89% accuracy increase, 35.44% faster inference, and 99.64% memory reduction compared to VGG11.
• Achieved 84.67% accuracy on Android in real-world settings, validating practical feasibility.

Conclusions:

• G-PPW-VGG11 is a feasible and efficient solution for practical wheat variety classification in intelligent agriculture.
• The model's lightweight design and high performance support mobile deployment and real-world applications.
• Public release of the model and dataset facilitates future research and development.