Precise identification and validation of microplastics using deep learning models in the Yellow River

Investigating the spatial distribution and seasonal variations of riverine microplastics (MPs) is vital to aquatic ecological research. Mass suspended sediment within the Yellow River introduces severe background noise, which impedes conventional deep learning models from accurately extracting fine edge features of microplastics. Based on water samples collected from the Zhengzhou reach of the Yellow River, this study systematically evaluates a matrix of 24 hybrid encoder-decoder segmentation architectures covering six backbones (ResNet and ResNeXt series) and four decoders (FPN, DeepLabV3+, UNet, UNet++). The objective is to identify the optimal single model configuration (including network depth, topological layout, and feature fusion strategy) capable of accurately extracting fine edge features of microplastics under severe sediment interference.. Beyond standard segmentation metrics, three dedicated indices including Counting Error (CE), Background Complexity (BC) and 95% Hausdorff Distance (HD95) are proposed to quantify the anti-noise robustness of each model. Grad-CAM visualization is adopted to interpret model spatial feature activation and target focus patterns, and micro-Fourier transform infrared (μ-FTIR) spectroscopy is utilized to cross-validate segmentation outputs. The ResNeXt-101 + UNet++ hybrid network delivers optimal performance. Equipped with grouped convolution and nested dense skip connections, this architecture bridges the information gap between shallow spatial details and deep semantic features, achieving a precision of 0.9534, a recall of 0.9991 and a mean absolute counting error (MAE) of 0.242. Microplastic abundances predicted by the model show excellent consistency with μ-FTIR measurements, addressing the common disconnection between image segmentation and spectroscopic identification in existing studies. This work develops an accurate intelligent pipeline for microplastic detection and quantification, facilitating routine surveillance and ecological risk assessment within the Yellow River Basin, while providing a methodological framework that warrants further validation in other sediment-laden aquatic environments.