Clasificarea automată a polenului folosind tehnici de învățare automată / Automatic pollen classification using deep learning techniques

Pollen is the main cause of seasonal allergies in humans. Millions of people suffer from such allergies and would benefit from better pollen monitoring and forecasts. While there are pollen monitoring networks in place, most of them rely on human technicians to do pollen counting and classification. More modern devices, that can operate without human intervention, have become more widespread but they still lack the intelligence or software to use the data that they can collect. This work focuses on developing fully automated pollen monitoring systems that use data from autonomous instruments. Data obtained from devices such as Rapid-E and BAA-500 devices are used to create new publicly available data sets of pollen species that are known allergens. These new data sets are vital when attempting to create intelligent systems that can replace humans in pollen monitoring tasks. These multi-modal data sets are used to build and train Machine Learning models capable of distinguishing pollen down to the plant species with an accuracy higher than that of human experts. This thesis is structured around the different types of measurements that can be made for pollen. In chapter 4 the data used is microscope imaging and the machine learning task is to segment and classify all the different particles present in the image. In chapter 5 the focus is on using multi-modal scattering and fluorescence imaging to classify pollen particles on data from Rapid-E devices. In the last results chapter 6, the task of predicting atmospheric pollen concentration is presented and some variables that impact the lifecycle of plants are discussed.