Improving Rice Classification with Deep Learning: A Study from Jilin Agricultural University

Strategic Insights -

Researchers at Jilin Agricultural University have made a breakthrough in rice classification using deep learning methods, according to a new study published in Current Plant Biology. The study aimed to develop a method for distinguishing rice kernels from different geographic environments, which is crucial for precision agriculture. The researchers proposed a method based on deep learning and image processing, which achieved an accuracy of 96.80% in recognizing rice from different origins.

Key Takeaways:

  • The researchers proposed a deep learning method to recognize rice from different origins, which achieved an accuracy of 96.80%.
  • The method used EfficientNet_b0 as the backbone network and introduced a dynamic adjustment strategy for the learning rate, removing the Dropout layer, and introducing grouped convolution to improve accuracy.
  • The study collected 30,000 images of Ji-Japonica 830 rice from ten different regions and used image segmentation and data enhancement to prepare the data for training and testing.
  • The researchers compared and tested four lightweight networks and four classical networks in the pre-training phase and found that EfficientNet_b0 obtained the highest accuracy of 93.38%.
  • The method performed well in terms of classification accuracy, parameters, time, and robustness, and could effectively distinguish rice kernels from different geographic environments.
  • The study highlights the potential of deep learning methods in precision agriculture and the importance of geographical traceability in rice classification.

Statistics:

  • The researchers collected a total of 30,000 images of Ji-Japonica 830 rice from ten different regions.
  • The method achieved an accuracy of 96.80% in recognizing rice from different origins.
  • The comparison of four lightweight networks and four classical networks showed that EfficientNet_b0 obtained the highest accuracy of 93.38% in the pre-training phase.

Sources:

  • "Improving EfficientNet_b0 for distinguishing rice from different origins: A deep learning method for geographical traceability in precision agriculture" Current Plant Biology, 2025, 43():100501.
  • http://www.journals.elsevier.com/current-plant-biology/
  • https://doi-org.sdpl.idm.oclc.org/10.1016/j.cpb.2025.100501