Neural Network Analysis for Microplastic Segmentation.

Area of Science:

• Environmental Science
• Computer Science
• Materials Science

Background:

• Microplastic contamination in beaches poses environmental challenges.
• Effective separation of microplastics from sand requires precise identification.
• Automated methods are needed for analyzing sand samples containing microplastics.

Purpose of the Study:

• To determine an optimal neural network for segmenting microplastic particles in beach sand images.
• To analyze the effectiveness of U-net and MultiResUNet architectures for tiny object detection.
• To develop computationally efficient models for microplastic identification.

Main Methods:

• Utilized a kernel weight histogram-based analytical process to evaluate neural network performance.
• Explored U-net and MultiResUNet architectures for microplastic segmentation.
• Developed reduced versions of U-net and MultiResUNet (Half U-net, Half MultiResUNet, Quarter MultiResUNet).
• Visualized network performance using TensorBoard.

Main Results:

• Initial encoder stages of U-net and MultiResUNet effectively capture small features.
• Later encoder stages were less effective for small feature detection.
• Half MultiResUNet achieved the best average recall-weighted F1 score (40%) and mIoU (26%).
• Reduced models required significantly fewer floating-point operations and parameters compared to original architectures.

Conclusions:

• Reduced U-net variants, particularly Half MultiResUNet, are highly effective for microplastic segmentation in sand.
• These optimized networks offer a computationally efficient solution for environmental monitoring.
• The findings facilitate improved methods for microplastic separation and analysis.