Deep learning-based automatic detection of productive tillers in rice

Abstract

The number of productive tillers per plant is one of the important agronomic traits associated with the grain yield of rice crop. However, manual counting of productive tillers is time-consuming, laborious and error-prone. In this study, a method for automatically detecting and counting productive tillers of rice crop was proposed based on deep learning convolutional neural network (CNN). The CNN model was trained using large amounts of in-field images taken by mobile phones from various varieties of rice crops under various environmental conditions. A Web app, integrating the trained CNN model and a Django server, was designed for fast and high-throughput detection of productive tillers. The performance of the Web app was evaluated for field-based practical applications. Results showed that the selected CNN model had a high precision and a fast detection rate. Through applying the Web app to 200 in-field images with 5 to 30 tillers per image, the number of productive tillers detected agreed well with manual counting data, regardless of rice variety or type of mobile phone used for image taking. The coefficients of determination between the Web app detection and manual counting of tillers were over 0.97 in all cases. Overall, compared to the manual counting, the accuracy of the Web app was over 99%. Furthermore, the performance of the Web app was not affected by the environmental conditions, such as illumination condition (cloudy or sunny) and water reflection in paddy fields.