HierbaNetV1: a novel feature extraction framework for deep learning-based weed identification

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Convolutional Neural Networks (CNNs) excel at feature extraction and pattern learning.
• Varying region of interest (ROI) sizes pose challenges in image analysis.
• Effective differentiation of crops from weeds is crucial for agricultural productivity.

Purpose of the Study:

• To propose HierbaNetV1, a novel CNN framework for robust feature extraction from images.
• To address the challenge of varying ROI sizes using diversified filters.
• To develop an accurate and generalizable model for crop-weed classification.

Main Methods:

• Developed HierbaNetV1, a feature extraction framework generating 3,872 feature maps per sample.
• Integrated low-level and high-level features for intensive and diversified learning.
• Created and utilized the SorghumWeedDataset_Classification for training and testing.

Main Results:

• HierbaNetV1 achieved a classification accuracy of 98.06% on the SorghumWeedDataset_Classification.
• The framework outperformed pre-trained models and state-of-the-art architectures.
• Ablation studies and component analysis confirmed HierbaNetV1's effectiveness and generalization across diverse datasets.

Conclusions:

• HierbaNetV1 offers a highly accurate and effective solution for crop-weed classification.
• The model demonstrates strong generalization capabilities, applicable to various crops and weeds.
• The research facilitates practical application through the HierbaApp and open-sourced code for community advancement.