Synthetic root images help AI learn when real annotations are scarce

Rhizotrons let researchers observe roots growing in soil through a transparent surface without digging up the plant. But turning those images into measurements remains difficult. Fine roots can resemble soil particles, boundaries can be faint, and overlapping structures complicate the task of marking which pixels belong to roots.

In a bioRxiv preprint posted on November 2, 2026, Fengjiao Lu and colleagues investigate whether synthetic images can help train root-segmentation models with fewer real annotations. The work has not been peer reviewed. Its central finding is conditional: synthetic pretraining helped most when real annotations were scarce, and the benefit depended on the model.

Building artificial roots on real soil

The researchers combined Stable Diffusion 1.5 with three complementary techniques. Low-Rank Adaptation, or LoRA, adapted the generator to the appearance of rhizotron roots. ControlNet guided the spatial arrangement of generated roots using edges derived from predefined root masks. Inpainting inserted the roots into real soil-background images.

Those masks came from existing annotations of Arabidopsis roots grown on agar plates. The team extracted individual root systems, rotated and rescaled them, and combined them into new arrangements. Each original mask was retained as the segmentation label for its generated image, avoiding the need to trace every synthetic example manually.

The resulting dataset contained 3,758 synthetic images with paired masks. This was not a pipeline independent of real material: it used real root photographs to adapt the generator, real soil backgrounds, and previously annotated root architectures.

Where pretraining helped

The authors compared three segmentation systems: nnU-Net v2, YOLO26 and Mask2Former. Each was trained using synthetic images alone, real annotations alone, or synthetic pretraining followed by fine-tuning on real annotations.

The complete real training set came from just two full-resolution rhizotron images, divided into 108 patches. Only 35 patches contained visible roots. A separate Arabidopsis test set comprised 15 full-resolution images.

Synthetic-only training performed worse than real-only training for all three models. Its highest Dice score was 0.596, achieved by YOLO26. Dice measures overlap between predicted root pixels and reference annotations; higher values indicate closer agreement.

With the complete real training set, synthetic pretraining raised nnU-Net v2’s Dice score from 0.786 to 0.799. It did not improve Dice for the other two models: YOLO26 declined slightly, while Mask2Former changed very little.

The stronger gains appeared in the lowest-annotation setting. Here, “10%” meant 10% of the training set’s annotated root pixels, not 10% of its images or annotation time. This subset contained nine patches, including two with visible roots. Synthetic pretraining raised nnU-Net v2’s Dice from 0.573 to 0.623 and Mask2Former’s from 0.460 to 0.624. YOLO26 showed only minor differences across annotation levels.

The team also tested transfer to eight unseen crops, including lettuce, maize, wheat and French bean, without additional crop-specific fine-tuning. Under limited real annotation, all three models improved on average across the cross-species evaluation set, which also included Arabidopsis as a reference. Gains became smaller with the complete training set and varied between species.

Human agreement is not a perfect target

The researchers separately asked seven annotators to label 10 visually clear patches and 10 ambiguous ones. They combined the annotations into probabilistic consensus maps and evaluated predictions using a confidence-weighted Soft Dice score. Pixels attracting disagreement contributed less than pixels with strong agreement.

On clear patches, nnU-Net v2 and Mask2Former approximately matched average human agreement. YOLO26 did not. On ambiguous patches, all three remained below the human average. These findings concern agreement within a small evaluation set, not a general demonstration of human-level root analysis.

Segmentation is only the first step

The authors identify several limits: one imaging platform, training annotations from one primary annotator, and synthetic architectures derived from agar-grown Arabidopsis. The current approach also addresses visible roots rather than reconstructing hidden portions behind soil.

Most importantly, the study did not test whether segmentation gains improve measurements such as root length or other root traits. The results support synthetic images as a useful supplement to scarce annotations, but validation across imaging conditions and against downstream biological measurements remains necessary.