Advanced multi-level ensemble learning with image-based features and FTIR data for accurate identification of micro- and nanoplastics

Reliable identification of micro- and nanoplastics (MNPs) remains difficult due to their complex morphologies, overlapping chemical signatures, and environmental variability. Conventional analytical techniques are often destructive, costly, and labor-intensive, limiting their suitability for standardized and real-time monitoring. Although recent artificial intelligence (AI)-assisted approaches have improved analytical throughput, they continue to exhibit critical limitations, including dataset dependence, noise sensitivity, and difficulty distinguishing chemically similar polymers. To address the limitations of both traditional analytical methods and current AI-based models, this study proposes a unified, robust, and interpretable MELP (Multi-level Ensemble Learning for micro-nanoPlastics) framework for accurate MNPs identification. MELP is a hierarchical architecture applied independently to image-derived morphological descriptors for MNPs and Fourier-transform infrared (FTIR) spectral signatures for microplastics (MPs), using a consistent model configuration across both modalities rather than a single unified multimodal network. In controlled imaging experiments, MELP surpassed XGBoost baseline, achieving 100% accuracy for clay–plastic discrimination and 99.8% for polymer-type classification under dry conditions, and maintaining 99.3% and 95.7% accuracy under wet conditions. Across FTIR benchmarks, MELP demonstrated superior generalizability, outperforming all state-of-the-art baselines, including deep neural networks (DNN), ensemble models, and the stacking-based MLStackXT framework, across accuracy, precision, recall, and Cohen’s kappa. Specifically, MELP achieved 96.9% accuracy on Kedzierski, 95.6% on Jung, and 76.0% on the more challenging Brignac dataset. Confusion matrix (CM) analyses confirmed a significant reduction in misclassification, achieving 100% accuracy in 9 of 12 MPs on Kedzierski, over 98% accuracy in 3 of 5 MPs on Jung, and 100% accuracy in multiple MPs on Brignac, with most others exceeding 88%. Interpretability analyses using SHapley Additive exPlanations (SHAP) and principal component analysis (PCA) further confirmed that MELP’s decisions are driven by chemically meaningful spectral, including RGB ratios under ultraviolet (UV) illumination and FTIR band assignments corresponding to polymer-specific functional groups rather than spurious correlations.