Convex Optimization of Environmental Processes

This study introduces a Kernel Ridge Regression (KRR)-based microplastic quantification method, enhanced by active learning strategies, to predict the number of microplastic particles from aggregate weight measurements. Our approach utilizes a subset of experimentally quantified samples to train the KRR model, which then predicts microplastic counts in remaining samples, thus reducing labor and resource use. The active learning strategies help in determining the most informative samples, thereby minimizing the number of necessary training samples without compromising prediction accuracy.Multiple experimental datasets are used to evaluate the performance of three active learning querying strategies: passive sampling, maximum variance reduction, and residual regression. Numerical results show that passive sampling significantly outperforms traditional leave-one-out cross-validation method by selecting highly representative samples for training. Moreover, we explore the influence of incorporating additional features on the prediction accuracy, finding that the effectiveness of these features varies depending on their correlation with microplastic counts across different datasets. The study demonstrates the potential of combining KRR with active learning to efficiently and effectively quantify microplastics, which is crucial for microplastic pollution monitoring and management.