Abstract:

Based on photos of rice collected hourly during the daytime and manually observed developmental stage data from Qiaoshang village, Luyang town, Zhongfang county, Huaihua city, Hunan province in 2023, a picture dataset containing developmental stages of rice which were seeding period, greening period, tillering period, jointing period, booting period, heading period, grain filling period as well as pre-seeding and the harvested period in a total of 10 periods had been established and processed using cutting、preprocessing and data enhancement techniques. Six representative deep learning network models, namely 18-layers residual network（ResNet18）, 50-layers residual network variant（ResNet50_vd）, lightweight convolutional network（MobileNetV3_large）, lightweight convolutional network（PPLCNet）, deep convolutional residual network（Xception41）, and densely connected convolutional network（DenseNet121）were selected, which were used as pre-training models to recognize the developmental stages of rice based on real pictures, the performances of the six models were compared in the training and test sets to evaluate their accuracy and loss rate to verify the feasibility of the deep learning model for the intelligent recognition of rice developmental stage, analyze its differences, and screen out the optimal rice developmental stage recognition model to promote its application in the business service. The results indicated that all models achieved a recognition accuracy of 92% or higher on the test set, among these models, Xception41 exhibited the highest recognition accuracy of 96.19%. The best model for recognition of pre-seeding, seeding, tillering, booting, grain filling and maturity periods of rice was Xception41, the best model for recognition of greening period of rice was ResNet50_vd, the best models for recognition of jointing and heading periods of rice were DenseNet121 and Xception41, and the best model for recognition of harvested period of rice were ResNet50_vd and DenseNet121 for test set. The study provided a new idea for the intelligent recognition of rice developmental stage, demonstrating the feasibility of deep learning models in recognition of rice live pictures and their potential to meet the demand of agricultural meteorological business services. 

Key words: Deep learning, Model, Accuracy, Developmental stage