Synthetic plant disease image generation to improve segmentation tasks in low-resource settings

Abstract

Accurate plant disease segmentation is often constrained by the availability of large, finely annotated datasets, particularly for rare diseases. This work presents a synthetic data generation pipeline that combines 3D leaf modelling with diffusion-based disease synthesis to address this limitation. Procedurally-generated leaf geometries are built in the 3D modelling package Blender to provide exact ground-truth masks, after which style-transfer is applied using Stable Diffusion, fine-tuned with Low-Rank Adaptation (LoRA) and guided by ControlNet conditioning to both preserve leaf structure and enforce correct lesion placement. The approach is evaluated on apple leaf diseases using a deliberately restricted subset of the PlantVillage dataset, simulating a controlled low-data-resource environment. Downstream task effectiveness is measured through leaf disease segmentation. The results show that combining data from the pipeline with limited real data leads to consistent improvements in segmentation performance.

Code and data availability

The authors publicly deposited the paper's annotated PlantVillage subset (75 images with segmentation masks) plus 300 synthetic images with ground-truth masks on Zenodo, directly reproducing this paper's phenotyping/segmentation data.