On the role of geometric representations for ultra-fine-grained crop variety identification
Abstract
Distinguishing plant varieties within a single species, an Ultra-Fine-Grained Visual Categorization (Ultra-FGVC) problem, remains largely unsolved under real-world agricultural conditions. While foundation models have significantly improved species-level plant identification, their ability to capture the subtle phenotypic variations that define cultivars remains unclear.In this work, we investigate which representation learning paradigm is best suited for infraspecific plant identification. We compare two approaches: (i) geometric self-supervised learning (DINOv2/v3), which learns representations by enforcing consistency across different sub-regions (patches) of the same image, and (ii) semantic alignment (BioCLIP 2), which aligns images with taxonomic textual descriptions. We further evaluate a hybrid strategy combining geometric pre-training with species-level supervised fine-tuning (Pl@ntNet), aiming to leverage the complementary strengths of both approaches. Experiments are conducted across three complementary datasets: a controlled market garden dataset (22 varieties), a participatory banana dataset (87 classes), and a grape variety dataset exhibiting a strong domain shift (34 varieties).
Our results show that geometric self-supervised representations provide the most suitable foundation for ultra-fine-grained recognition, consistently outperforming semantic alignment approaches. While species-level supervision yields modest improvements, its contribution remains limited compared to the gains obtained from geometric pre-training, indicating that the discriminative signal at the varietal level is primarily morphological rather than semantic. Furthermore, domain-specific adaptation provides small but consistent gains, and naïve multi-organ aggregation can dilute discriminative signals in asymmetric taxa such as Musa (bananas).
These findings demonstrate that ultra-fine-grained crop variety identification relies primarily on geometric sensitivity, and highlight the importance of designing models and aggregation strategies that preserve fine-scale morphological information for agrobiodiversity monitoring.
Related
- Human similarity judgments reveal fine-grained representations of dynamic social scenes
- Vision Transformers versus convolutional neural networks for fine-grained orchid genus identification in a species-rich, data-poor flora: a controlled benchmark on the Orchidaceae of New Guinea
- Geometric reshaping of task-relevant representations in the primary visual cortex supports perceptual decisions
- StackHPpred: A Stacking-based ensemble learning framework for the identification of peptide hormones using multi-view feature representations
- Variety-Specific Responses of Mulberry to Copper-Induced Stress
BibTeX
@article{tibi2026role,
title = {{On the role of geometric representations for ultra-fine-grained crop variety identification}},
author = {Charlotte Tibi and Ilyass Moummad and Alexis Joly and Yves Caraglio and Christophe Jenny and Pierre Bonnet and Hervé Goëau},
journal = {Smart Agricultural Technology},
year = {2026},
month = sep,
doi = {10.1016/j.atech.2026.102442},
url = {https://inria.hal.science/hal-05736639v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - On the role of geometric representations for ultra-fine-grained crop variety identification
AU - Charlotte Tibi
AU - Ilyass Moummad
AU - Alexis Joly
AU - Yves Caraglio
AU - Christophe Jenny
AU - Pierre Bonnet
AU - Hervé Goëau
PY - 2026
DA - 2026/09/03
JO - Smart Agricultural Technology
DO - 10.1016/j.atech.2026.102442
UR - https://inria.hal.science/hal-05736639v1
AB - Distinguishing plant varieties within a single species, an Ultra-Fine-Grained Visual Categorization (Ultra-FGVC) problem, remains largely unsolved under real-world agricultural conditions. While foundation models have significantly improved species-level plant identification, their ability to capture the subtle phenotypic variations that define cultivars remains unclear.In this work, we investigate which representation learning paradigm is best suited for infraspecific plant identification. We compare two approaches: (i) geometric self-supervised learning (DINOv2/v3), which learns representations by enforcing consistency across different sub-regions (patches) of the same image, and (ii) semantic alignment (BioCLIP 2), which aligns images with taxonomic textual descriptions. We further evaluate a hybrid strategy combining geometric pre-training with species-level supervised fine-tuning (Pl@ntNet), aiming to leverage the complementary strengths of both approaches. Experiments are conducted across three complementary datasets: a controlled market garden dataset (22 varieties), a participatory banana dataset (87 classes), and a grape variety dataset exhibiting a strong domain shift (34 varieties). Our results show that geometric self-supervised representations provide the most suitable foundation for ultra-fine-grained recognition, consistently outperforming semantic alignment approaches. While species-level supervision yields modest improvements, its contribution remains limited compared to the gains obtained from geometric pre-training, indicating that the discriminative signal at the varietal level is primarily morphological rather than semantic. Furthermore, domain-specific adaptation provides small but consistent gains, and naïve multi-organ aggregation can dilute discriminative signals in asymmetric taxa such as Musa (bananas). These findings demonstrate that ultra-fine-grained crop variety identification relies primarily on geometric sensitivity, and highlight the importance of designing models and aggregation strategies that preserve fine-scale morphological information for agrobiodiversity monitoring.
SN - 2772-3755
ER -