RiceNet: A two stage machine learning method for rice disease identification

Rice diseases are one of the key factors affecting the yield and quality of rice and their identification is critical. However, taking rice disease images in the field is time-consuming and laborious. Furthermore, the images taken in the rice field contain complex background information, such as weeds, soil, and unwanted sections of rice plants. The use of only limited annotated images to rapidly and accurately identify diseases in complex scenes is key problem to be solved. A two-stage method called RiceNet was proposed to identify important four rice diseases, including rice panicle neck blast, rice false smut, rice leaf blast, and rice stem blast. In the first instance, YoloX was used to detect the diseased parts of rice, and the original rice disease images were clipped to form a new rice disease patch dataset according to the detection results. In the second stage, Siamese Network was used to identify the rice disease patch dataset obtained in the first stage. For the detection stage, the mAP of YoloX for rice disease images was 95.58%, and the comparison experiment showed that YoloX achieved the highest detection performance. For the identification stage, Siamese Network achieved the identification accuracy of 99.03%, which is higher than that of other models. The experimental results show that the proposed RiceNet model was superior to state-of-the-art methods. Moreover, it achieved a high detection speed with the smallest weight size for the identification of rice diseases.