Using deep learning for pruning region detection and plant organ segmentation in dormant spur-pruned grapevines

Area of Science:

• Agricultural Robotics
• Computer Vision
• Viticulture

Background:

• Vineyard management costs persist due to labor-intensive selective operations like winter pruning.
• Robotic solutions are emerging in agriculture, but their application to grapevines remains under-explored.

Purpose of the Study:

• To fine-tune and test deep neural networks for detecting pruning regions (PRs) and segmenting grapevine organs (GOs).
• To develop components for an autonomous robotic system for grapevine winter pruning.

Main Methods:

• Fine-tuning Faster R-CNN with 1215 RGB images for PR detection, tested on 232 images.
• Training Mask R-CNN with 119 RGB images for GO segmentation (cordon, arm, spur, cane, node), tested on 60 images.

Main Results:

• Faster R-CNN achieved high PR detection rates, particularly for visible spurs (0.97).
• Mask R-CNN demonstrated effective GO segmentation, with nodes being best segmented (0.88).
• Shoot-thinned (ST) grapevines showed higher segmentation recall (0.85) than control (0.80), indicating canopy management's impact.

Conclusions:

• Deep learning models show promise for automating grapevine winter pruning tasks.
• Visibility significantly impacts pruning region detection accuracy.
• Canopy management strategies can improve the performance of AI-based robotic solutions in viticulture.