Fusion of NIR and LIBS spectra via interpolation for plastic classification

The excessive use and non-degradability of plastics aggravate pollution and resource depletion, making rapid and accurate classification essential for recycling. This study combines laser-induced breakdown spectroscopy (LIBS) and near-infrared spectroscopy (NIR) to exploit their complementary atomic and molecular information for plastic classification. However, the significant resolution disparity between the two spectroscopic modalities results in data scale mismatch, which limits the effectiveness of multimodal data fusion. To address this issue, cubic spline interpolation is employed to resample the NIR spectra, matching its data scale with LIBS and preserving spectral continuity, thereby enhancing fusion accuracy. Based on this resampling strategy, two fusion schemes, low-level and mid-level, are developed. The mid-level fusion approach incorporates Variable Importance in Projection (VIP) and Variable Importance (VI) methods for feature selection to extract key discriminative features. The fused features are subsequently classified using architectures based on Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models. This study implements a dual plastic classification strategy based on both type and color. Specifically, black plastics are identified using LIBS, while non-black plastics are classified using LIBS and NIR. For plastics of same type but different colors, NIR spectroscopy is employed for color differentiation. For non-black plastic classification. Experiments include four fusion strategies and thirteen classification models, with performance evaluated through five-fold cross-validation. The results show that the improved mid-level data fusion based on cubic spline interpolation achieved the best classification performance, with a prediction accuracy of 99.2 %. Compared with the unimodal LIBS and NIR models, the proposed fusion method improved classification accuracy by 19.29 % and 4.76 %, respectively.