Image classification of forage grasses on Etuoke Banner using edge autoencoder network

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate forage identification is crucial for intelligent livestock breeding, impacting animal growth, product quality, and health.
• Traditional manual methods for forage identification are inefficient, inaccurate, and disrupt natural livestock behavior.
• Automated forage recognition systems are needed to support grassland evaluation, management, and precision feeding strategies.

Purpose of the Study:

• To develop an optimized Convolution Neural Network (CNN) algorithm, the Edge Autoencoder Network (E-A-Net), for accurate forage species identification.
• To establish a comprehensive forage grass dataset for the Etuoke Banner region, facilitating research and development in this area.
• To improve the accuracy of herbage recognition, especially in complex natural environments, by addressing limitations of existing methods.

Main Methods:

• Constructed a novel dataset of 3889 forage grass images across 22 categories from Etuoke Banner.
• Implemented data preprocessing with random cutout enhancement and background removal via threshold-based image segmentation.
• Developed the E-A-Net by integrating a Sobel operator for edge information extraction and a pre-trained autoencoder for a hard attention mechanism, fusing multi-scale and overall features.

Main Results:

• The E-A-Net algorithm demonstrated significant improvements in forage species identification accuracy.
• Data preprocessing techniques, including background removal, enhanced recognition in complex environments.
• The proposed E-A-Net outperformed benchmark models (VGG16, ResNet50, EfficientNetB0) with improved f1-scores of 1.6%, 2.8%, and 3.7% respectively.

Conclusions:

• The E-A-Net effectively addresses the loss of edge and overall feature information in deep convolutional networks.
• This automated forage identification method provides a robust foundation for ecological assessments and precision livestock management.
• The developed E-A-Net offers a significant advancement over traditional methods and basic CNNs for forage recognition applications.