Text guidance is powerful but prompt-sensitive for weakly-supervised leaf symptom segmentation

Abstract

Accurate segmentation of plant disease symptoms is essential for crop monitoring and phenotyping, yet it typically requires costly pixel-level annotations. Weakly supervised semantic segmentation (WSSS) alleviates this burden using image-level labels, but its performance depends on the quality of spatial priors such as class activation maps (CAMs). We investigate whether text-guided segmentation with the Segment Anything Model 3 (SAM3) can serve as an alternative weak supervision signal. Three pseudo-mask generation strategies are compared: (i) CAMs refined with SAM or SAM3, (ii) zero-shot text-guided SAM3, and (iii) a hybrid approach combining weak spatial cues with text prompts. The resulting pseudo-masks are used to train a DeepLabV3 model. Text guidance alone matches or outperforms conventional WSSS, achieving up to 0.46 IoU without spatial supervision and 0.61 IoU on a public dataset, although performance is sensitive to text prompt formulation. The hybrid strategy improves robustness, reaching 0.50 IoU on the primary dataset and 0.58 IoU on the additional dataset while reducing prompt sensitivity. Overall, text guidance is a promising alternative to conventional weak supervision, while hybrid approaches provide a more robust solution for plant disease segmentation.

Plant phenotyping relevance

植物病害症状の画像セグメンテーション手法を開発・比較し、病斑の疑似マスク生成とセグメンテーション性能を評価しているため、植物フェノタイピング手法が中心です。

Code and data availability

The paper's own oilseed rape leaf disease image dataset (2,540 RGB images, 601 pixel-level annotations) is explicitly made publicly available on Recherche Data Gouv under DOI 10.57745/R3HZOQ, which is an allowed URL. A companion dataset describing the foliar pathogens (DOI 10.57745/UCEKI8) is also public. The pixel-ann