Harnessing AI for non-destructive in situ root imaging and phenotyping

Roots are essential for plant growth, but traditional methods of studying roots are resource-intensive and harmful. With advances in image processing techniques, innovative methods for in situ root studies have emerged, providing non-destructive root imaging.

However, ground shading in images is the current challenge, leading to fragmented root systems and a loss of structural integrity. And this fragmentation hinders the accurate assessment of root phenotypes. Although deep learning approaches such as SegRoot and ChronoRoot have improved root image recognition, issues such as root breakage and soil cover still remain.

Advances in image restoration, especially in situ root identification, are crucial for accurate root phenotype assessment. Techniques such as with techniques such as generative adversarial networks (GANs) show potential in this part, but still require refinement.

In July 2023, Plant phenomics published a research article titled “Application of improved UNet and EnlightenGAN for segmentation and reconstruction of in situ roots.” In this study, researchers proposed using EnlightenGAN for root reconstruction by manipulating the light intensity in targeted areas.

The team previously developed the RhizoPot platform, which can collect the complete root images non-destructively. Early stages showed accurate segmentation of roots with DeeplabV3+. However, there were inaccuracies in the analysis of root diameter and surface area. Ongoing research has improved the accuracy of in situ root segmentation, but small pieces covered by soil still remain unidentified.

By comparing the deep learning models UNet, SegNet and DeeplabV3+ on an original root dataset, the study found that DeeplabV3+ (Xception) had the best overall performance. However, each model had its strengths and weaknesses in root identification. Ablation experiments with various enhancements on UNet showed increased performance in both mIOU and F1 scores, suggesting that these modifications successfully addressed the limitations of the models.

Transfer learning with the improved UNet on the reconstructed root dataset demonstrates good versatility and robustness. EnlightenGAN was used for root generation, with each iteration progressively improving root reconstruction. Phenotypic parameters were analyzed using RhizoVision Explorer software, which revealed a significant correlation with actual values. However, the reconstructed roots resulted in changes in root length and surface area.

The study performed a thorough model comparison, which highlighted the DeeplabV3+’s capabilities, but also noted the limitations of the model in recognizing major roots. The improved UNet was chosen for root segmentation because of its scalability and potential for future improvements. Finally, the study proposed various combinations of UNet and EnlightenGAN for different purposes, ranging from accurate segmentation to dataset expansion and unsupervised training.

Overall, the study shows a significant advance in root reconstruction technology, which is a new approach to root phenotyping analysis.