A Point-Cloud Segmentation Network Based on SqueezeNet and Time Series for Plants

Area of Science:

• Plant Science
• Computer Vision
• Agricultural Technology

Background:

• Accurate plant phenotyping is crucial for understanding genetic traits and advancing precision agriculture.
• Segmenting 3D point clouds of plant organs is fundamental for extracting phenotypic parameters.

Purpose of the Study:

• To develop a novel point-cloud downsampling method addressing sample imbalance issues.
• To architect a deep learning framework for plant point cloud segmentation using SqueezeNet principles.
• To enhance segmentation accuracy by incorporating time-series data as input variables.

Main Methods:

• A novel point-cloud downsampling technique was introduced.
• A SqueezeNet-based deep learning framework was developed for semantic segmentation.
• Time-series data were integrated to improve network performance.
• The MeanShift algorithm was applied for instance segmentation based on semantic segmentation results.

Main Results:

• Semantic segmentation achieved high accuracy for maize (99.35% Precision, 99.26% Recall, 99.30% F1-score, 98.61% IoU) and tomato (97.98% Precision, 97.92% Recall, 97.95% F1-score, 95.98% IoU).
• Instance segmentation accuracy reached 98.45% for maize and 96.12% for tomato.
• The integration of time-series data significantly improved segmentation accuracy.

Conclusions:

• The developed method effectively segments plant point clouds, offering a robust solution for plant phenotypic extraction.
• This research contributes to advancements in ideotype selection and precision agriculture.
• The novel downsampling and deep learning framework show significant potential for agricultural applications.