A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Environmental Engineering and Management

Thesis (M. Sc.) - Asian Institute of Technology, 2026

The problem of pollution with microplastic debris has become one of the most pressing ones because of the resistance of plastics to biodegradation and existing challenges of the classical approach to their rapid detection in aquatic ecosystems. This work studied the potential for applying the method of Electrical Sensing Zone (ESZ) combined with machine learning and deep learning techniques for classifying particles. The chosen classes include three types of microplastics: polyethylene terephthalate (PET), polypropylene (PP), and polystyrene (PS); two classes of inorganic particles: silicon dioxide (SiO₂) and calcium carbonate (CaCO₃); and Chlorella algae as a class of biological particles.The raw signals have been generated under laboratory conditions as voltage-time waveforms. Further preprocessing comprised denoising, baseline correction, peak detection, and pulse extraction. The signal of Chlorella algae particles was close to the baseline. Thus, the main classification analysis involved microplastic particles and inorganic particles' signals. Pulse signals have been encoded both as engineered feature vectors for feature-based machine learning and as fixed-length segments (windows) for CNN classification. The strategies considered included the following: five-class machine learning, binary machine learning, binary CNN classification, flat three-particle CNN classification, step-wise hierarchical three-particle CNN classification, and PET/PS/PP subtype CNN classification.The performance of five-class feature-based machine learning models proved low since the best results were obtained by Random Forest with a baseline test accuracy of 44.13%, followed by Boosting (42.76%) and SVM (38.43%). For the binary machine learning classifier, which could classify microplastics against inorganic particles, the performance was not taken into account here. CNN models demonstrated stronger classification capabilities: binary CNN yielded 91.98% of test accuracy; flat three-particle CNN {u2013} 58.13%; step-wise hierarchical three-particle CNN {u2013} 74.47%; PET/PS/PP subtypes CNN {u2013} 51.05%.