This repository presents a novel approach for improving the identification of pollen particles using weakly supervised domain adaptation. The method is designed to work well even with only a small amount of labeled data. By combining transfer learning, specifically domain adaptation, nested cross-validation, and uncertainty analysis, the approach provides more accurate and reliable pollen classification across different imaging devices.
The results and method are published here:
https://link.springer.com/article/10.1007/s10489-024-06021-9
Our approach builds on a CNN trained to classify pollen types from airborne measurement data by integrating weakly supervised domain adaptation. We first train a baseline multi-modal CNN with a labelled dataset for pollen classification. Then, to bridge the gap between controlled labelled datasets and real-world operational data, we retrain the model using expert-verified measurements from manual pollen counts. This domain adaptation step leverages large amounts of unlabeled operational data while minimizing the discrepancy between automatic predictions and manual standards. By fine-tuning the classifier, the adapted model improves the correlation with manual measurements by 23% and reduces prediction uncertainty by 38% compared to the baseline model.
Runs nested cross-validation on a CNN model.
This script helps evaluate the model in a robust way and selects the best hyperparameters.
Trains the final model using all available data and the hyperparameters selected during cross-validation.
Retrains the previously trained model by adding more data (often treated as ground truth).
This step helps fine-tune and improve the model’s performance. More information is available in the paper ola2023 (found in the repository).
Calculates correlation distributions to measure uncertainty in both the model predictions and the ground truth.
This provides insights into how reliable the classifications are. Details can be found in the stoten2022 paper (included in the repository).
All scripts use helper functions and modules stored in the Libraries/ folder.
- Works well with small labeled datasets
- Can be used across different devices
- Improves accuracy through transfer learning and retraining
- Provides uncertainty estimates for better interpretation of results
- Fully reproducible workflow for scientific use
If you use this work in your research, please cite: Matavulj, P., Jelic, S., Severdija, D. et al. Domain adaptation for improving automatic airborne pollen classification with expert-verified measurements. Appl Intell 55, 430 (2025). https://doi.org/10.1007/s10489-024-06021-9
Contributions and suggestions are welcome.
Feel free to open an issue or submit a pull request.