# Efficient and scalable training set generation for automated pollen monitoring with Hirst-type samplers

## TL;DR

This paper introduces an automated pipeline for generating training datasets for pollen detection, enabling scalable and cost-effective AI-based monitoring.

## Contribution

A novel pipeline combining one-shot detection and refinement to generate high-quality annotations with minimal manual effort.

## Key findings

- ViT-based models outperformed ResNet50 in classification tasks across datasets.
- ResNet50-based models achieved highest accuracy for detecting Ambrosia pollen in real-world conditions.
- Cross-dataset generalization remains challenging, highlighting the need for domain adaptation techniques.

## Abstract

Automated pollen detection is essential for ecological monitoring, allergy forecasting, and biodiversity research. However, existing methods rely heavily on manual or semi-automated annotations, limiting scalability and broader applicability. We introduce a highly automated training dataset generation pipeline that combines one-shot detection with systematic refinement, producing tens of thousands of high-quality annotations from bright-field microscopy while significantly reducing manual effort and annotation costs. Using multi-regional datasets from France, Hungary, and Sweden, we trained object detection models on seven pollen taxa and evaluated their performance on both external pure and mixed species slides and real-world airborne samples. We assessed the reusability of pretrained vision models for pollen detection, aiming to reduce the need for extensive retraining. Using linear probing, we identified foundational Vision Transformers (ViTs) as the most effective feature extractors and integrated them into Faster R-CNN detection models. We benchmarked these models against ResNet50, a widely adopted backbone in biological imaging. On held-out regions of the training datasets, our models achieved high performance in both classification and detection tasks. On independent reference slides from other datasets, ViTs continued to outperform ResNet50 in classification. However, in full object detection and under real deployment conditions, ResNet50-based models remained competitive and achieved the highest accuracy for detecting Ambrosia, a major allergen with public health significance. Cross-dataset generalization remains a challenge, underscoring the need for domain adaptation techniques such as stain normalization and data augmentation. This study establishes a scalable framework for AI-assisted pollen monitoring, supporting large-scale slide digitization and enabling applications in long-term ecological research, allergen surveillance, and automated biodiversity assessment.