Source-linked AI summary
Meta-Learning for Data-Efficient Plant Growth Estimation via Vision Transformers and Fuzzy Clustering
Sheikh Hasan Elahi, Rusith Chamara Hathurusinghe Dewage, Habib Ullah, Muhammad Salman Siddiqui, Rakibul Islam, Fadi Al Machot
TL;DR
Plant growth estimation requires efficient learning when greenhouse labels are costly and scarce. The paper combines ViT embeddings, fuzzy c-means task construction, and gradient-based meta-learning, finding that structured tasks and second-order methods support reliable few-shot regression while intra-cluster selection has limited, dataset-dependent effects.
Problem
Greenhouse plant growth estimation needs timely monitoring, but manual phenotyping is labor-intensive and deep learning methods typically require large labeled datasets.
Method
The framework uses ViT feature embeddings, fuzzy c-means clustering to construct structured few-shot tasks, and gradient-based meta-learning.
Results
Across two plant datasets, MAML++ achieved the best overall few-shot performance, while second-order methods outperformed first-order Reptile and classical baselines.
Takeaways & Limitations
Task construction in embedding space is a primary performance driver, whereas intra-cluster support selection has limited and dataset-dependent impact.
Abstract
from arXiv · showhide
Accurate plant growth estimation is essential for greenhouse monitoring, yet obtaining labeled data remains costly and time-consuming. To address this, we propose a few-shot regression framework that combines Vision Transformer (ViT) feature embeddings, clustering-based task construction, and gradient-based meta-learning, and show that task construction in embedding space is a primary driver of performance. The approach leverages an unlabeled image pool to organize data into structured tasks using fuzzy c-means clustering, enabling efficient learning from a small number of labeled samples. We systematically evaluate meta-learning methods and show that second-order methods (e.g., Model-Agnostic Meta-Learning variants such as MAML++) outperform classical baselines in the few-shot regime. Furthermore, intra-cluster support selection has a limited and dataset-dependent impact. Experiments on two plant datasets show that structured task design combined with meta-learning enables reliable plant growth estimation under severe label scarcity.
I. INTRODUCTION
Plant growth estimation supports timely greenhouse management, but manual phenotyping and label-intensive deep learning limit high-frequency monitoring. The paper addresses this gap with few-shot meta-learning and embedding-space task construction.
- Few-shot learning and meta-learning target plant phenotyping settings where labeled data are scarce and manual measurements are labor-intensive.Traditional measurements are reliable but unsuitable for high-frequency monitoring, while deep learning typically requires large labeled datasets.
- The proposed framework combines Vision Transformer embeddings, fuzzy c-means representative sampling, and gradient-based meta-learning for few-shot plant growth regression.
- Clustering-based task construction organizes few-shot regression tasks in embedding space, while the study evaluates cross-dataset generalization on cucumber and lettuce.
- The study reports that task construction dominates sampling heuristics and that second-order meta-learning is important for plant growth regression.
- Structured sampling combined with meta-learning is presented as a practical solution for greenhouse phenotyping systems.
B. Meta-Learning for Low-Data Adaptation
The paper situates its approach at the intersection of meta-learning, Vision Transformer representations, and embedding-based sampling for agricultural few-shot tasks. It addresses the underexplored integration of representative sampling and meta-learning across plant species.
- B. Meta-Learning for Low-Data Adaptation: Meta-learning seeks rapid adaptation to new tasks from limited support samples, with MAML extensions including Meta-SGD and MAML++.
- C. Vision Transformers and Embedding-Based Sampling: Vision Transformers provide plant-structure representations and their embeddings can support sample selection through clustering.
- C. Vision Transformers and Embedding-Based Sampling: Fuzzy c-means clustering is presented as a soft sampling approach suited to capturing continuous variation in plant morphology.
- C. Vision Transformers and Embedding-Based Sampling: The integration of representative sampling and meta-learning within MAML remains underexplored across multiple plant species.
- The evaluation includes a cucumber dataset collected under controlled greenhouse conditions across 29 plants and 16 days of imaging.
B. Competition Lettuce Dataset
The lettuce-dataset evaluation uses ViT embeddings, fuzzy c-means clustering, and gradient-based meta-learning to construct localized few-shot regression tasks. The dataset contains 388 annotated top-view lettuce images with canopy-width labels, while the proposed pipeline compares multiple meta-learning variants under a common setup.
- B. Competition Lettuce Dataset: 388 top-view RGB lettuce images with canopy-width annotations are split into 310 training and 78 test images, producing 1,240 augmented training samples.The dataset represents a rosette structure distinct from cucumber.
- B. Competition Lettuce Dataset: The pipeline resizes plant images and extracts 768-dimensional embeddings with a pretrained ViT-Base/16 feature extractor.The embeddings are used to represent global morphological structure.
- B. Competition Lettuce Dataset: Fuzzy c-means organizes embeddings into morphologically coherent regions and supplies membership confidence for representative support selection.The number of clusters is selected on a validation set to avoid test-set leakage.
- B. Competition Lettuce Dataset: Each cluster defines a localized task with a representative support set and a query set sampled from remaining points assigned to that cluster.This construction is intended to expose the model to growth-pattern variation across plant structures.
- B. Competition Lettuce Dataset: The regression model is a two-layer MLP trained through task-based adaptation, with MAML, Meta-SGD, and MAML++ providing alternative gradient-based meta-learning procedures.MAML uses a fixed learning rate, Meta-SGD learns per-parameter rates, and MAML++ adds per-step learning rates and a multi-step meta-loss.
- B. Competition Lettuce Dataset: All methods share the same architecture and task construction, with validation-selected hyperparameters, multiple random seeds, and an 80:20 train/test split.Meta-training uses batches of tasks sampled from the cluster distribution.
V. RESULTS
The evaluation uses held-out test sets over 20 random seeds, with standardized growth labels during training and millimeter-scale RMSE reporting. These choices define the protocol for comparing methods.
- V. RESULTS: 20 random seeds are used to evaluate all methods on held-out test sets.
- V. RESULTS: Growth labels are standardized during training and de-standardized for evaluation.
- V. RESULTS: All RMSE values are reported in millimeters.
A. Main Performance Comparison
MAML++ achieves the strongest few-shot performance across cucumber and lettuce, reducing error toward full-data baselines while showing stable convergence across seeds.
- The comparison uses ViT features, FCM-based task construction, representative sampling, and MAML-based optimization against few-shot and full-data baselines.
- 35.13 ± 1.55 mm: MAML++ improves over MAML on cucumber from 42.89 ± 1.70 mm, an 18% reduction, and outperforms Meta-SGD by over 6 mm.
- MAML++ reduces the gap to full-data performance using fewer than 6% of the augmented training set, with results within roughly 6–13 mm of the best full-data baseline.
- MAML++ consistently achieves the lowest RMSE across 20 seeds and shows lower variance, particularly on cucumber.
B. Cluster Selection
Validation-based cluster selection identifies moderate granularity as effective for constructing few-shot tasks, with K = 7 performing best on both datasets.
- K = 7 yields the lowest validation RMSE on both datasets, while performance remains stable between K = 7 and K = 10.
- Smaller cluster counts, including K = 3 and K = 5, generally produce higher error than the moderate-granularity settings.
- Cluster count is selected exclusively on a training-derived validation split, avoiding test-set leakage and supporting reproducible evaluation.
C. Ablation of Within-Cluster Support Selection Strategies
Support-set coverage across clusters matters more than the specific representative-selection heuristic used within each cluster.
- Balanced sampling enforces equal cluster coverage, whereas unbalanced sampling selects globally high-membership samples without cluster-level constraints.
- 27.74 vs. 29.41 mm: random sampling outperforms representative sampling on lettuce, while cucumber differences remain below 1 mm.
- Intra-cluster selection shows no consistent representative-sampling advantage when the number of samples per cluster is fixed.
- The results indicate that support construction is primarily driven by cluster-level coverage, with intra-cluster heuristics playing a secondary role.
D. Statistical Significance
Statistical tests show that MAML systematically outperforms Reptile, while MAML++ offers a significant advantage over MAML on cucumber but not lettuce.
- p < 0.001, W = 0: MAML significantly outperforms Reptile on both datasets, indicating a systematic performance gap.
- MAML++ significantly improves over MAML on cucumber, but their lettuce difference is not significant (p = 0.546).
VI. DISCUSSION
The discussion finds that second-order meta-learning methods outperform first-order Reptile for few-shot plant growth regression, while MAML++ gains over MAML and representative sampling depend on the dataset.
- Second-order methods show a large, statistically significant performance gap over first-order Reptile across both datasets.The discussion attributes this pattern to more precise parameter adaptation through inner-loop differentiation.
- MAML++ achieves the best overall performance, but its advantage over MAML is substantial on cucumber and negligible on lettuce.The authors relate this dataset dependence to differences in dataset variability.
- Balanced sampling improves stability, whereas FCM-based representative sampling does not consistently outperform random selection.The reported pattern suggests that embedding-space coverage is more important than prototypicality.
- MAML++ reaches competitive performance using less than 6% of the augmented training data, with consistent trends across cucumber and lettuce.This supports the framework's relevance to settings with limited annotations and differing plant structures or imaging conditions.
VII. CONCLUSION
The paper presents a data-efficient framework combining ViT representations, FCM-based task construction, and gradient-based meta-learning for few-shot plant growth estimation. Across cucumber and lettuce, MAML++ achieves the best overall performance, while structured representative sampling offers limited, dataset-dependent benefits.
- MAML++ achieves the best overall few-shot performance, reaching 35.13 mm RMSE on cucumber and 29.13 mm on lettuce.
- The framework combines ViT-based representations, FCM-based task construction, and gradient-based meta-learning.
- The experiments show that second-order meta-learning is effective for low-data plant growth regression, while representative sampling benefits are limited and dataset-dependent.
- The study is limited to controlled greenhouse imagery and single-trait regression, with future work targeting broader crop diversity, temporal modeling, and uncertainty-aware prediction.