Investigation of the data fusion of spectral and textural data from hyperspectral imaging for the near geographical origin discrimination of wolfberries using 2D-CNN algorithms

Abstract

Deep convolutional neural networks have been applied to hyperspectral imaging (HSI) and have significantly improved modelling performance in many spectral analysis tasks due to their automatic extraction of relevant features. Using visible and near infrared hyperspectral (Vis-NIR) data, two-dimensional convolutional neural network (2D-CNN) discrimination models between the spectra of wolfberries and their corresponding classes of geographical origins were established and optimized using various variable selection and data fusion methods. The interval variable iterative space shrinking analysis (iVISSA), the uninformative variable elimination (UVE) algorithm, competitive adaptive reweighted sampling (CARS) and the iterative retained information variable (IRIV) algorithms were used to extract the feature wavelengths and compare the modelling effects; and then the 72 optimal wavelengths were extracted by the iVISSA algorithm. To extract the textural features of images, grey-level co-occurrence matrix (GLCM) analysis was conducted on the first principal component image. Models using variable selection methods based on low-level fusion data were superior to the corresponding methods based on single spectral data. The model based on iVISSA achieved the best result on mid-level fusion, the prediction set accuracy and mean F1 were 97.34% and 100%, respectively. Finally, optimized models of spectral-textural data were employed to identify the geographical origins of wolfberries. In general, the results showed that 2D-CNN model combined with fusion data of spectral and textural information can obtain excellent identification effect for the near geographical origins of wolfberries. This study may help develop an online detection system of near geographical origins of wolfberries.