Domain adaptation with unlabeled data for model transferability between airborne particle identifiers

As the most common causes of seasonal allergies, pollen affects approximately 30% of the world population. The proper information on the number of airborne allergens can significantly reduce its negative health and economic impact. For this reason, there is a growing network of automatic airborne particle monitors deployed. However, the calibration of such devices is a tedious task. Developing a deep learning classifier may allow model transferability between the devices. To investigate this approach, we employed data from two Rapid-E particle identifier devices, in a multi-class pollen identification task. We aim to improve the performance of models trained with data from one device and tested on another device. To our knowledge, this is the first attempt to apply any domain adaptation technique with unlabeled data between automatic airborne particle identifiers. Convolutional Neural Networks were constructed with two outputs to simultaneously perform pollen identification and domain adaptation. A simple gradient reversal layer between the domain classifier and the feature extractor promotes the emergence of not just discriminative features related to the classification task but also features invariant to the domain shifts in data. The development of a method for model transferability has a huge practical value for pollen monitoring since it reduces the costs of collecting labeled data