High-Throughput in situ Root Image Segmentation Based on the Improved DeepLabv3+ Method.

Area of Science:

• Plant Science
• Computer Vision
• Agricultural Technology

Background:

• Rhizotrons method is crucial for studying plant root growth phenotypes.
• Manual segmentation of root images is inefficient and error-prone, hindering research.
• Automated segmentation is needed to advance root system analysis.

Purpose of the Study:

• To develop and validate an automated image segmentation method for plant root systems.
• To improve the DeepLabv3+ convolutional neural network (CNN) architecture for root image segmentation.
• To compare the novel method's performance against traditional manual and U-net segmentation techniques.

Main Methods:

• Utilized an improved DeepLabv3+ CNN architecture for root image segmentation.
• Trained and validated the model using in situ micro-root window images of cotton roots.
• Assessed segmentation performance using F1-score, recall, precision, and Spearman rank correlation.

Main Results:

• Achieved high segmentation performance with F1-score (0.9773), recall (0.9847), and precision (0.9702) after 80 training epochs.
• Demonstrated strong correlation (Spearman's rho = 0.9667, r² = 0.9449) between automated and manual root length measurements.
• The improved DeepLabv3+ method showed superior accuracy in complex soil environments compared to manual and U-net methods.

Conclusions:

• The improved DeepLabv3+ CNN offers an accurate and efficient solution for segmenting plant root systems from micro-root images.
• This automated approach significantly overcomes the limitations of traditional manual segmentation methods.
• The method is particularly advantageous for root system analysis in homogeneous soil environments.