01 — Overview
POLLEN13K is the first large-scale pollen grain image dataset, comprising more than 13,000 labeled objects extracted from aerobiological microscope images. It was built to support the development and evaluation of automatic pollen grain classification systems using modern machine learning and deep learning techniques.
Aerobiology — the study of airborne biological particles — plays a critical role in medicine, biology, and agronomy. Reliable estimation of airborne pollen abundance informs allergy risk assessment, plant phenology monitoring, and crop yield forecasting. However, manual analysis by qualified experts is time-consuming and limits scalability. POLLEN13K addresses this bottleneck by providing a curated, expert-validated resource for training and benchmarking automated classifiers.
02 — Dataset Classes
Under the guidance of aerobiology experts, all segmented objects were manually assigned to one of five categories. The Cupressaceae class contains very few examples (43 objects) and was excluded from the baseline classification experiments due to data imbalance.
| Label | Class | Objects | Notes |
|---|---|---|---|
| 1 | Corylus avellana (well-developed) | 1,850 | Normal hazel pollen grains |
| 2 | Corylus avellana (anomalous) | 903 | Morphologically aberrant hazel grains |
| 3 | Alnus (well-developed) | 9,558 | Alder pollen — dominant class |
| 4 | Debris | 999 | Air bubbles, dust, fungal spores, non-pollen objects |
| 5 | Cupressaceae | 43 | Cypress family pollen (low sample count) |
Each object is provided as a color RGB image (84×84 px), together with its corresponding binary segmentation mask and a version with a uniform green background — facilitating experiments with different input modalities.
03 — Comparison with Existing Datasets
POLLEN13K substantially surpasses previous publicly available pollen datasets in terms of the number of annotated objects, making it uniquely suited for training data-hungry deep learning models.
| Dataset | Objects | Image Type | Resolution |
|---|---|---|---|
| Duller's Pollen Dataset | 630 | Grayscale | 25×25 |
| POLEN23E | 805 | Color | ≥250 px/dim |
| Ranzato et al. | 3,686 | Color | 1024×1024 (multi-grain) |
| POLLEN13K (Ours) | >13,000 | Color | 84×84 |
04 — Data Acquisition & Segmentation
Airborne samples were collected on adhesive tapes mounted on a rotating drum (moving at 2 mm h⁻¹ under a suction hole). Pollen grains were inspected on daily segments using a Leitz Diaplan bright-field microscope with a 5 MP CMOS sensor. Pollen walls were stained with a mounting medium containing basic fuchsin.
Acquired images contained heavy background noise from debris, dust, fungal spores, and air bubbles introduced during mounting. A dedicated segmentation pipeline was developed to reliably locate and extract individual objects.
05 — Baseline Classification Results
To establish baseline performance, several standard machine learning classifiers were evaluated using HOG and LBP feature descriptors, followed by experiments with deep convolutional neural networks. The dataset was split 85% / 15% for training and testing, using stratified sampling to preserve class proportions. The weighted F1 score was chosen as the primary metric to account for class imbalance.
Among classical approaches, RBF SVM with HOG features achieved the best result (accuracy: 86.6%, F1: 0.857). Among deep learning models, AlexNet trained on an augmented dataset (combining green-background and noisy-background images) reached an average F1 score of 0.87. SmallerVGGNet achieved 0.85 on the same augmented set. The most common misclassification occurred between well-developed Corylus avellana and Alnus grains, which share similar surface textures.
06 — Citation
07 — Acknowledgements
This research was carried out in collaboration with Ferrero HCo, which financed the project and enabled the collection of aerobiological samples from hazelnut plantations. The authors thank the aerobiology experts who contributed to sample collection, labelling, and validation.