Heterogeneity classification based on hyperspectral transmission imaging and multivariate data analysis

Abstract

Hyperspectral images contain numerous spectral bands, which can provide rich information for image classification. Based on this, this paper combines multivariate data analysis to realize the heterogeneity classification in hyperspectral transmission images. Firstly, the hyperspectral image acquisition system is built. The microcontroller is used to control the stepper motor driver, which further controls the cross sliding platform to scan the XY two-dimensional plane, while the spectrometer is synchronously triggered externally to acquire hyperspectral images during the scanning process. Secondly, the acquired hyperspectral images are processed, including filtering and denoising, extracting regions of interest, coding to obtain labeled heterogeneity samples, and dividing the sample set. Finally, nine bands are selected by the successive projection algorithm, and three discriminative models of partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and k-nearest neighbor (kNN) are used to discriminate the types of heterogeneities. The classification results show that the SVM model has the best overall classification (the highest classification accuracy is 98.83%), PLS-DA is the second-best (the highest classification accuracy is 97.5%), and kNN has the worst prediction effect (the highest accuracy is only 83.33%) on different sample set division algorithms. In addition, the determination coefficients R2 of both PLS-DA and SVM models are greater than 0.97 with different sample set division algorithms, and the root mean square error (RMSE) is less than 0.16, which effectively proves the robustness of the models.