A Wavelet-Enhanced Dual-Path Multiscale Residual Attention Network for Real-Time Denoised Classification and Segmentation of Marine Microplastics in Hyperspectral Imagery

Abstract

The study addresses the urgent need for accurate and efficient detection of microplastics in marine environments, which pose a significant threat to aquatic ecosystem health. The goal is to develop a robust deep learning framework capable of classifying and segmenting microplastics in hyperspectral imagery (HSI). A novel Wavelet-Enhanced Dual-Path Multiscale Residual Attention Network (WED-MRAN) is proposed. The model integrates wavelet-based denoising, dual-path feature extraction, multiscale processing, and residual attention mechanisms to improve spectral and spatial feature representation in hyperspectral data. The method is evaluated using a specialized HSI dataset with pixel-level annotations for microplastic types (PE, PP, PS) and natural backgrounds (seawater, sand, algae). The WED-MRAN model achieved a classification accuracy of 92.5%, IoU of 88.2%, and Dice Similarity of 89.0%. Additionally, it demonstrated a low RMSE of 0.032 in terms of denoising performance. It outperformed traditional architectures such as CNN, U-Net, and ResNet, as well as other models incorporating wavelet or attention mechanisms. This work introduces a unique combination of wavelet denoising and advanced dual-path residual attention networks for hyperspectral microplastic detection. The model's superior performance in noisy data conditions highlights its potential for real-time, practical deployment in marine environmental monitoring applications.