Researchers from the Hefei Institutes of Physical Science (HFIPS) of the Chinese Academy of Sciences have developed a new algorithm in the direction of near-infrared spectroscopy (NIRS) that is suitable for high-throughput identification of the authenticity of crop varieties. Results were published in Infrared Physics & Technology.

The authenticity of crop varieties is of great significance in variety protection and seed breeding. Traditional methods for authentic identification of crop varieties, such as DNA molecular identification, isoenzyme identification, and field identification, have the disadvantages of complicated operations, time-consuming, samples damage, environmental pollution, and lagging detection results. Therefore, an effective method is urgently needed to realize crop variety authenticity identification. 

As a rapid detection technology, NIRS is environmental-friendly, highly-sensitive, and non-destructive.

In this study, the self-developed high-throughput seed quality sorting instrument based on NIRS has achieved rapid identification and sorting of individual seeds.

Based on this instrument, the researchers proposed an improved convolution neural network—the InResSpectra network, to help achieve more accurate seed variety identification. This was an optimized Inception network. It removed the 1 × 1 convolution branch to reduce the complexity of the model, and increased the residual element of the ResNet network, which accelerated the training of the neural network and improved accuracy.

The researchers applied the developed system and the InResSpectra network for the identification of 24 wheat varieties and 21 rice varieties, and achieved 95.35% and 93.07% accuracy, respectively, providing an effective method for the spectroscopic identification of the authenticity of crop varieties.

This work was supported by the National Natural Science Foundation of China, the Anhui Provincial Science and Technology Major Project, and the Anhui Provincial Key Research and Development Program.

 Heat maps of the confusion matrices of wheat (a) and rice (b) sample sets identified by InResSpectra. (Image by XU Zhuopin)