In Situ Root Dataset Expansion Strategy Based on an Improved CycleGAN Generator

Area of Science:

• Plant Biology
• Computer Vision
• Agricultural Science

Background:

• In situ root research is essential for understanding plant water and nutrient uptake.
• Deep learning for root segmentation requires extensive labeled datasets, which are often scarce.
• Existing segmentation methods struggle with background variations in root images.

Purpose of the Study:

• To develop a novel method for augmenting in situ root image datasets.
• To improve the accuracy and stability of deep-learning-based root segmentation.
• To enhance the versatility of datasets for plant root analysis.

Main Methods:

• Utilized an improved CycleGAN generator for dataset expansion.
• Implemented a spatial-coordinate-based method for target-background separation.
• Employed time-division soil image acquisition for diverse culture medium integration.

Main Results:

• The augmentation strategy significantly improved segmentation performance (e.g., 0.63% increase in mIoU on validation, 33.6% on generalization).
• The proposed method demonstrated superior speed, accuracy, and stability over traditional thresholding.
• Enhanced dataset versatility through diverse culture medium integration.

Conclusions:

• The proposed dataset augmentation strategy is feasible and practical for in situ root research.
• The improved CycleGAN and segmentation methods enhance the analysis of root phenotypes and dynamics.
• Future work includes advanced rendering for shading simulation and broader dataset creation.