Hyperspectral imaging for classification of bulk grain samples with deep convolutional neural networks

Abstract

Near Infrared hyperspectral imaging (HSI) offers a fast and non-destructive method for seed quality assessment through combining spectroscopy and imaging. Recently, convolutional neural networks (CNN) have shown to be promising tools for red-green-blue (RGB) image or spectral cereal classification. This paper describes the design and implementation of deep CNN models capable of utilizing both the spatial and spectral dimension of HSI data simultaneously for analysis of bulk grain samples with densely packed kernels. Classification of eight grain samples, including six different wheat varieties, were used as a test case. The study shows that the CNN architecture ResNet, originally designed for RGB images, can be adapted to use the full spatio-spectral dimension of the HSI data through adding a linear down sample layer prior to the conventional ResNet architecture. Using traditional spectral pre-processing methods before passing the data to the CNN does not improve the classification accuracy of the networks, while a channel-wise image standardization improves the accuracy significantly. The modified ResNet applied to the full spatio-spectral dimension has a classification accuracy of up to 99.75 ± 0.02%, outperforming both purely spectral (86.5 ± 0.1%) and purely spatial (98.70 ± 0.01%) based methods in terms of accuracy, indicating that utilizing spatio-spectral correlation can improve sample classification, but also that grain classification is primarily solved using spatial information. The findings reported in this paper demonstrate how CNN networks can be designed to leverage spatio-spectral information in hyperspectral data. The combination of HSI and spatio-spectral CNN networks shows a possible method for fast prediction of bulk grain quality parameters where both spectral and spatial properties of the grains are important.