Assessing the potential of deep learning to distinguish wild and domesticated grapevine pollen
Abstract
Discriminating the pollen signature of wild Vitis vinifera ssp. sylvestris and domesticated grapevines V. vinifera ssp. vinifera can help reconstruct past vegetation and human practices in the Euro-Mediterranean region and Western Asia. The study evaluates several methods relying on automated image analysis powered by Convolutional Neural Networks (CNNs) to discriminate these two subspecies using pollen. Its primary objective is to classify the tricolporate morphotype between the two subspecies, while complementary analyses focus on both tricolporate and inaperturate morphotypes. Pollen from 14 wild and 12 domesticated grapevine individuals from the Vassal-Montpellier ampelographic collection (France) was processed and digitized using light microscopy. A detection CNN was applied to generate thousands of pollen images. A classification CNN was then used to discriminate between the tricolporate fertile pollen of the two subspecies from a subset of 20 individuals. Models were trained and evaluated using either pollen from the same individuals, or pollen from different individuals. Classification accuracy reached 64.7 ± 1.4% in the first design but dropped to 51.7 ± 5.5% when evaluated on unseen individuals. These results suggest that models did not find consistent differences between subspecies, but rather captured features specific to the individuals themselves, which may reflect greater inter-individual than inter-subspecies variation in pollen morphology or size. Using pollen images from the 26 individuals, we found no differences in pollen size or production between the subspecies or the tricolporate and inaperturate morphotypes. More advanced microscopy techniques may help confirm whether differences between the two subspecies are only very subtle or absent.
Related
- Assessing the feasibility of up-scaling grapevine physiological trait estimation using multispectral UAV imagery
- Sodium nitrophenolates as a potential tool for the management of grapevine trunk diseases
- Changes to task relevance in redundant motor environments distinguish structure learning from adaptation
- Low-cost rhizotron imaging and zero-shot deep-learning resolve temporal, spatial, and genetic variation in grapevine rootstock root systems
- Advancing grapevine phenology predictions: a comparison of regression-based statistical learning approach and conventional models in mediterranean environment.
BibTeX
@article{gimenez2026assessing,
title = {{Assessing the potential of deep learning to distinguish wild and domesticated grapevine pollen}},
author = {Betty Gimenez and Céline Devaux and Sébastien Joannin and Odile Peyron and Laurent Bouby and Pranabendu Misra and Sourish Das},
journal = {Review of Palaeobotany and Palynology},
year = {2026},
month = sep,
doi = {10.1016/j.revpalbo.2026.105720},
url = {https://hal.science/hal-05749858v1},
note = {CC BY-NC},
}
RIS
TY - JOUR
TI - Assessing the potential of deep learning to distinguish wild and domesticated grapevine pollen
AU - Betty Gimenez
AU - Céline Devaux
AU - Sébastien Joannin
AU - Odile Peyron
AU - Laurent Bouby
AU - Pranabendu Misra
AU - Sourish Das
PY - 2026
DA - 2026/09/15
JO - Review of Palaeobotany and Palynology
DO - 10.1016/j.revpalbo.2026.105720
UR - https://hal.science/hal-05749858v1
AB - Discriminating the pollen signature of wild Vitis vinifera ssp. sylvestris and domesticated grapevines V. vinifera ssp. vinifera can help reconstruct past vegetation and human practices in the Euro-Mediterranean region and Western Asia. The study evaluates several methods relying on automated image analysis powered by Convolutional Neural Networks (CNNs) to discriminate these two subspecies using pollen. Its primary objective is to classify the tricolporate morphotype between the two subspecies, while complementary analyses focus on both tricolporate and inaperturate morphotypes. Pollen from 14 wild and 12 domesticated grapevine individuals from the Vassal-Montpellier ampelographic collection (France) was processed and digitized using light microscopy. A detection CNN was applied to generate thousands of pollen images. A classification CNN was then used to discriminate between the tricolporate fertile pollen of the two subspecies from a subset of 20 individuals. Models were trained and evaluated using either pollen from the same individuals, or pollen from different individuals. Classification accuracy reached 64.7 ± 1.4% in the first design but dropped to 51.7 ± 5.5% when evaluated on unseen individuals. These results suggest that models did not find consistent differences between subspecies, but rather captured features specific to the individuals themselves, which may reflect greater inter-individual than inter-subspecies variation in pollen morphology or size. Using pollen images from the 26 individuals, we found no differences in pollen size or production between the subspecies or the tricolporate and inaperturate morphotypes. More advanced microscopy techniques may help confirm whether differences between the two subspecies are only very subtle or absent.
SN - 0034-6667
ER -