Deep Learning: Individual Maize Segmentation From Terrestrial Lidar Data Using Faster R-CNN and Regional Growth...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Light Detection and Ranging (Lidar) offers precise 3D data acquisition for plant phenotyping.
• Individual plant segmentation is essential for high-throughput phenotyping but remains challenging.
• Deep learning excels in object detection and segmentation tasks.

Purpose of the Study:

• To develop and validate a method for segmenting individual maize plants from terrestrial Lidar data.
• To leverage deep learning and regional growth algorithms for accurate maize segmentation.
• To assess the method's performance across varying planting densities.

Main Methods:

• A hybrid approach combining Faster R-CNN (region-based convolutional neural network) for stem detection and regional growth algorithms for plant reconstruction.
• 3D Lidar point clouds were processed into deep images for training the Faster R-CNN model.
• The trained model identified maize stems, serving as seeds for the regional growth algorithm.

Main Results:

• The combined deep learning and regional growth method achieved high segmentation accuracy (r, p, F > 0.9) across different planting densities.
• Segmented maize plant heights showed a strong correlation with manually measured heights (R² > 0.9).
• The approach demonstrated robust performance in segmenting individual maize plants.

Conclusions:

• The proposed method effectively addresses the challenge of individual maize segmentation using terrestrial Lidar data.
• This technique shows significant potential for advancing automated, high-throughput phenotyping in precision agriculture.
• Deep learning combined with regional growth offers a powerful solution for extracting detailed 3D crop traits.