IFCNN-based fusion of GAF and MTF encoded near-infrared spectral images for quantitative analysis of microplastics

Microplastics pollution has become the second-largest global environmental issue. Therefore, accurate and rapid quantification of microplastics is of great significance. This study proposed a novel quantitative approach that integrated near-infrared (NIR) spectral image transformation with deep learning-based image fusion. NIR spectra were collected from mixtures of five types of microplastics and sand at six concentration levels. Gramian Angular Fields (GAF) and Markov Transition Fields (MTF) were employed to convert one- dimensional spectral data into two-dimensional feature images, which were subsequently fused using the Improved Fusion Convolutional Neural Network (IFCNN). A two-dimensional Convolutional Neural Network (2D-CNN) model was developed based on the fused GAF-MTF images for concentration prediction. Comparative analysis with conventional models proved that the proposed GAF-MTF-IFCNN method has transcended conventional quantitative models of NIR-PLS, NIR-SVR and image fusion method of GASF-GADF-IFCNN. The prediction results of NIR-PLS, NIR-SVR and GASF-GADF-IFCNN models prove that the models' performance have been greatly improved by fusing GAF and MTF images. Furthermore, GAF-MTF-IFCNN model has achieved the optimal R values of 0.956, 0.765, 0.986, 0.976, 0.999 for PE, PET, PP, PS, PVC, respectively. This work demonstrated the potential of the GAF and MTF encoding combined with IFCNN-based image fusion in NIR spectroscopy to be used as a rapid, accurate, and reliable method for the quantitative detection of various types of microplastics.