A novel semi-supervised framework for UAV based crop/weed classification.

Area of Science:

• Agricultural Science
• Computer Science
• Environmental Science

Background:

• Excessive agrochemical use for weed control poses environmental and agronomic risks.
• Precision agriculture (PA) and smart farming require accurate weed identification for targeted control.
• Current supervised classification systems for weed detection using Unmanned Aerial Vehicle (UAV) imagery are labor-intensive due to the need for extensive labeled data.

Purpose of the Study:

• To develop an optimized semi-supervised learning approach for classifying crops and weeds at early growth stages.
• To reduce the reliance on large labeled datasets in UAV-based weed detection systems.
• To improve the efficiency and accuracy of weed classification for precision agriculture applications.

Main Methods:

• Development of a semi-supervised generative adversarial network (GAN) for crop and weed classification.
• Utilizing a generator within the GAN to create additional training data for the discriminator.
• Employing Red Green Blue (RGB) images captured by a quadcopter in pea and strawberry fields.

Main Results:

• The proposed semi-supervised GAN achieved an average accuracy of 90% with 80% unlabeled training data.
• The system demonstrated superior performance compared to standard supervised learning classifiers.
• The method proved effective for weed classification even with limited labeled samples and reduced training time.

Conclusions:

• The developed semi-supervised GAN is a viable and efficient technique for crop and weed classification using UAV imagery.
• This approach significantly addresses the challenge of limited labeled data in precision agriculture.
• The system offers a cost-effective and time-saving solution for targeted weed management, contributing to sustainable farming practices.