Improving EfficientNet_b0 for distinguishing rice from different origins: A deep learning method for geographical traceability in precision agriculture

Rice is one of the important crops for food supply, and there are multiple differences in the quality of rice grown in different geographic environments, which have an important impact on subsequent yield, economic efficiency, and food processing. Most of the current computer vision-based rice kernel classification focuses only on different varieties. In this study, we propose a method based on deep learning and image processing to recognize rice from different origins. First, Ji-Japonica 830 rice was collected from ten different regions, and a total of 30,000 images were obtained through image segmentation and data enhancement to participate in the training and testing of the model. Four lightweight networks and four classical networks were compared and tested in the pre-training phase, where EfficientNet_b0 obtained the highest accuracy of 93.38%, and then EfficientNet_b0 was improved by introducing a dynamic adjustment strategy for the learning rate, removing the Dropout layer, and introducing a grouped convolution, which resulted in 96.80% accuracy. The experimental results show that the method performs well in terms of classification accuracy, parameters, time, and robustness, and can effectively distinguish rice kernels from different geographic environments. • The same variety of rice grown in different areas was collected. • Four lightweight pre-trained networks were chosen. • EfficientNet_b0 has been improved from three perspectives. • The improved algorithm was validated on three different types of images. • A dynamic learning rate adjustment strategy was proposed.