Roots, integral to plant growth, have long been challenging to study due to resource-intensive and damaging traditional methods. The advent of innovative in situ root imaging techniques, propelled by advancements in image processing, has introduced non-destructive approaches to root studies. However, a persistent challenge in these studies is the issue of soil shading in images, resulting in fragmented root systems and compromised structural integrity. Despite the strides made by deep learning techniques like SegRoot and ChronoRoot, obstacles such as root breakage and incomplete soil coverage persist.
Addressing the need for improved root phenotype assessment, researchers have explored advanced image restoration techniques, with a particular focus on in situ root identification. While approaches involving generative adversarial networks (GANs) show promise, refinement is still required.
In a pivotal research article published by Plant Phenomics in July 2023, titled “Application of Improved UNet and EnlightenGAN for Segmentation and Reconstruction of in Situ Roots,” a novel method for root reconstruction using EnlightenGAN and manipulation of light intensity was proposed. The research team, building on their earlier development of the RhizoPot platform for nondestructive root image collection, aimed to overcome previous challenges, particularly inaccuracies in root diameter and surface area analysis.
Comparative analysis of deep-learning models UNet, SegNet, and DeeplabV3+ on an original root dataset revealed that DeeplabV3+ (Xception) demonstrated superior overall performance. However, each model exhibited specific strengths and weaknesses in root identification. Ablation experiments with various improvements on UNet showcased increased performance, addressing the models’ limitations effectively.
Transfer learning with the enhanced UNet on the reconstructed root dataset displayed versatility and robustness. The study utilized EnlightenGAN for root generation, progressively enhancing root reconstruction with each iteration. Phenotypic parameters were analyzed using RhizoVision Explorer software, indicating a significant correlation with actual values. Notably, reconstructed roots showed variations in root length and surface area.
While the study underscored DeeplabV3+’s capabilities, it also acknowledged limitations in recognizing main roots. The enhanced UNet, chosen for its scalability and potential for future enhancements, was designated for root segmentation. The research proposed diverse combinations of UNet and EnlightenGAN for various purposes, ranging from precise segmentation to dataset expansion and unsupervised training.
In conclusion, this study marks a substantial leap forward in root reconstruction technology, introducing a novel approach to root phenotyping analysis. The findings hold promise for transforming how researchers and scientists explore the intricate world of plant roots, providing new avenues for accurate and non-destructive root studies.