Classification of rice varieties using hyperspectral imaging with multi-dimensional fusion convolutional neural networks

Suitable variety of seeds planted under suitable conditions can achieve stable grain quality and yield. Therefore, accurately identifying the variety of seeds is very meaningful. In this study, near-infrared hyperspectral imaging was used to classify six varieties of rice seeds with deep learning. Based on the one-dimensional (1D) spectra, two-dimensional (2D) images, and three-dimensional (3D) hyperspectral images, the corresponding single convolutional neural network (CNN) models for classifying rice seed varieties were proposed. In addition to the single CNN models, multi-dimensional decision fusion neural network (MDDF-CNN) and multi-dimensional feature fusion convolutional neural network (MDFF-CNN) were proposed using the 1D spectra and 2D images. The overall results showed 1D CNN obtained better performance than 2D CNN and 3D CNN. Compared to 1D CNN, MDDF-CNN and MDFF-CNN have improved overall classification accuracy by 2.43 % and 3.97 %, with classification accuracy exceeding 90 %. The results illustrated the feasibility of using a multi-dimensional CNN to identify rice seed varieties, and the fusion of 1D spectra and 2D images showed great potential for better classification performance. The results would help to explore the feasibility of developing the combined use of low-cost spectrometers and cameras under the limited budget for rice seed variety identification and other applications.