Deep learning-based microplastic identification using FTIR spectra

Name: Thar, Aeint Shune

keyword: Convolutional neural networks (CNNs)

; Deep learning (DL)

; Desktop application

; Fourier-transform infrared (FTIR) spectroscopy

; Fusion

; Linear dimensionality reduction (LDR)

; Machine learning (ML)

; Membrane filters

; Microplastics

Abstract: The escalating concern of microplastic pollution in the environment, particularly in aquatic ecosystems, calls for advanced and reliable detection methodologies. This thesis presents a comprehensive study in three key areas: comparing machine learning (ML) and deep learning (DL) techniques, optimizing model performance through linear dimensionality reduction and fusion methods, and developing a desktop application for real-time microplastic classification using Fourier-transform infrared (FTIR) spectroscopy. In the first part, the study compares various ML and DL models to ascertain their effectiveness in classifying microplastics amid spectral noise challenges, particularly those arising from membrane filters in FTIR spectra. A detailed evaluation of traditional ML algorithms such as Support Vector Classification (SVC) and K-Nearest Neighbors (KNN), alongside Convolutional Neural Networks (CNNs) like LeNet5, is conducted. The analysis reveals that while ML models showcase high accuracy, the DL model LeNet5 outperforms them, achieving an accuracy of 96.93\%, thereby highlighting the advanced pattern recognition capabilities of CNNs in this domain. The second part delves into optimizing the performance of DL models for microplastics spectral classification by implementing linear dimensionality reduction and fusion techniques. Five linear dimensionality reduction methods, including Principal Component Analysis (PCA), Factor Analysis (FA), Fast Independent Component Analysis (Fast ICA), Non-negative Matrix Factorization (NMF), and Truncated Singular Value Decomposition (Truncated SVD), are explored for their efficacy in conjunction with CNN architectures. Furthermore, the study examines the impact of four fusion strategies: Shallow Feature Fusion (Data Fusion), Deep Feature Fusion, Hard Decision Fusion, and Soft Decision Fusion. The findings indicate that Soft Decision Fusion markedly enhances classification accuracy, emphasizing its potential in improving microplastics classification accuracy. In the final part, this research culminates in the development of a desktop application that integrates the optimized DL models and five linear dimensionality reduction techniques. This application offers a practical, real-time solution for microplastics classification, making it accessible to a wider range of users in environmental monitoring. It serves as a valuable tool for deploying developed models in real-world scenarios to detect and classify microplastics efficiently. This thesis demonstrates the potential of AI-driven models for efficient microplastics classification using FTIR spectroscopy, providing a scalable solution for environmental monitoring. The findings advocate for further refinement and adoption of these technologies to support global efforts in microplastics pollution mitigation.

Thammasat University. Thammasat University Library

Address: BANGKOK

Email: preserv@tu.ac.th

Name: Seksan Laitrakun

Role: advisor

Created: 2024

Modified: 2568-05-15

Issued: 2025-04-21

วิทยานิพนธ์/Thesis

application/pdf

DOI: 10.14457/TU.the.2024.90

eng

DegreeName: Master of Engineering

Level: Master's Degree

Descipline: Artificial Intelligence and Internet of Things

Grantor: Thammasat University

©copyrights Thammasat University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	Thar.pdf	80.41 MB

ใช้เวลา

0.031193 วินาที