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 by relying on cheaper forms of annotation, such as image-level labels, bounding boxes, or point annotations, to produce supervision (pseudo-mask) for training a segmentation model. We investigate whether text-guided segmentation with the Segment Anything Model 3 (SAM3) can serve as an alternative as it directly segments regions matching a text prompt description of the disease symptoms.Three pseudo-mask generation strategies are compared: (i) class activation maps (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 fully supervised model (DeepLabV3+). 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.
Related
- A hybrid supervised learning model for the classification of a low-volume EEG dataset employing powerful ensembles
- Characterization of the knock-in AppSAA mouse model as a powerful tool for Alzheimer's disease preclinical research.
- Does (co-)variation between boldness and growth differ between sexes? A nuanced and powerful evaluation of the pace of life hypothesis.
- FastGxC: Fast and powerful context-specific eQTL mapping in bulk and single-cell data
- Bioprocessing of Talaromyces purpurogenus for production of dihydromonascorubrin with a powerful antioxidant and antimicrobial activities integrated with the network-guided target prediction and molecular docking analyses.
BibTeX
@article{dubois2026text,
title = {{Text guidance is powerful but prompt-sensitive for weakly-supervised leaf symptom segmentation}},
author = {Romane Dubois and Lydia Bousset and Stéphane Jumel and Melen Leclerc and Nicolas Parisey and Alexis Joly},
journal = {Smart Agricultural Technology},
year = {2026},
month = jul,
doi = {10.1016/j.atech.2026.102544},
url = {https://hal.science/hal-05692432v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Text guidance is powerful but prompt-sensitive for weakly-supervised leaf symptom segmentation
AU - Romane Dubois
AU - Lydia Bousset
AU - Stéphane Jumel
AU - Melen Leclerc
AU - Nicolas Parisey
AU - Alexis Joly
PY - 2026
DA - 2026/07/10
JO - Smart Agricultural Technology
DO - 10.1016/j.atech.2026.102544
UR - https://hal.science/hal-05692432v1
AB - 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 by relying on cheaper forms of annotation, such as image-level labels, bounding boxes, or point annotations, to produce supervision (pseudo-mask) for training a segmentation model. We investigate whether text-guided segmentation with the Segment Anything Model 3 (SAM3) can serve as an alternative as it directly segments regions matching a text prompt description of the disease symptoms.Three pseudo-mask generation strategies are compared: (i) class activation maps (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 fully supervised model (DeepLabV3+). 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.
SN - 2772-3755
ER -