Behind the Leaves: Estimation of Occluded Grapevine Berries With Conditional Generative Adversarial Networks

Area of Science:

• Viticulture and Oenology
• Computer Vision
• Artificial Intelligence

Background:

• Accurate grape yield estimation is vital for the competitive global wine market.
• Berry counting offers a promising, non-destructive, and automatable method for harvest prediction.
• Occlusion by leaves presents a significant challenge to accurate berry counting.

Purpose of the Study:

• To develop an advanced method for estimating grape yield by accurately counting berries.
• To address the challenge of berry occlusion by leaves using deep learning.
• To improve the accuracy of non-destructive, automated harvest estimation techniques.

Main Methods:

• Utilized generative adversarial networks (GANs), a deep learning approach.
• Trained the GANs on images of non-occluded berries to learn patterns.
• Developed a method to generate probable scenarios behind occluding leaves.

Main Results:

• The proposed method significantly improved the accuracy of berry count estimates.
• The estimates were closer to manually counted references compared to traditional methods.
• The approach adapts to local vineyard conditions by analyzing visible berries.

Conclusions:

• Generative adversarial networks offer a powerful solution for estimating grape yield by overcoming berry occlusion.
• This deep learning method enhances the reliability of non-destructive harvest prediction in viticulture.
• The technique can identify areas needing augmentation without explicit hidden data input.