Deep Learning Based Greenhouse Image Segmentation and Shoot Phenotyping (DeepShoot).

Area of Science:

• Plant Science
• Computer Vision
• Bioinformatics

Background:

• High-throughput plant phenotyping demands automated analysis of large image datasets.
• Variability in plant appearance and experimental setups necessitates advanced machine and deep learning for plant structure segmentation.
• Automated detection and segmentation of plant structures in complex optical scenes are crucial for plant research.

Purpose of the Study:

• To introduce DeepShoot, a GUI-based software tool for automated segmentation and quantitative analysis of greenhouse-grown plant shoots.
• To leverage pre-trained U-net deep learning models for efficient and accurate plant image analysis.
• To provide a user-friendly solution for researchers without advanced IT skills.

Main Methods:

• Development of a GUI-based software tool named DeepShoot.
• Utilization of pre-trained U-net deep learning models trained on Arabidopsis, maize, and wheat.
• Application of the tool to segment plant shoots from various rotational side- and top-view images.

Main Results:

• The algorithmic framework achieved an average accuracy of over 90% for automated segmentation of plant shoots.
• DeepShoot demonstrated superior performance in both precision and processing time compared to shallow, conventional, and encoder backbone networks.
• The tool successfully segmented images acquired at different developmental stages and from diverse phenotyping facilities.

Conclusions:

• DeepShoot provides an efficient solution for automated segmentation and phenotypic characterization of greenhouse-grown plant shoots.
• The software is suitable for end-users lacking advanced IT expertise.
• While trained on specific species, DeepShoot can be applied to other plant species with similar optical properties.