In this paper, motivated by CNN combining global features of MLP to minimize the attention on the scene, we propose a parallel architecture model. Specifically, the model combines CNN\'s excellent inductive bias ability with MLP\'s salient ability to obtain the global features. Simultaneously, a dataset of 45547 diseases crop images of natural scenes has been gathered through collecting and cleaning. Extensive experiments have been conducted on the above dataset, demonstrating the superiority of our proposed model over other models. Under the condition of maintaining lower parameters, the accuracy rate is up to 92.5%.