Soybean gray spot disease is one of the most common disease of soybean, which may affect soybean yield seriously. In order to reduce the economic loss caused by this disease and achieve a precision pesticide, it is necessary to grade the different level of gray spot disease. However, due to the similarities between different levels of disease, the automated grading of this disease is still a challenge. This paper proposes a deep voting model to grade different levels of soybean gray spot disease by weighted integrating some basic convolutional neural network (CNN) models with genetic algorithm. First, some basic CNN models are trained by transfer learning, and then three models with the highest accuracy are chosen to be integrated, with the optimal weighted parameters learned by genetic algorithm. The proposed model was trained and evaluated on a private dataset with 32000 images of four levels of soybean gray spot disease. The experimental results show that the grading accuracy of proposed model reaches 94.9% on the test set with 4800 images, which is about 4% higher than the accuracy of the basic CNN model. This study will facilitate the real-application of automated disease grading in precision agriculture.