Citrus is widely cultivated in China and has many excellent varieties. There are many excellent varieties of citrus which are widely cultivated in China. However, citrus is susceptible to pest and disease infections during growth, which seriously affects the yield and quality of citrus. Common diseases include ulcer disease, deficiency disease and soot disease, etc. Insect pests include red spider and leaf miner moth, etc. Drug pests include herbicides and acaricides. The development of citrus industry is closely related to the control of diseases and insect pests. In order to realize the rapid and accurate identification of diseases and insect pests on citrus leaves, an Att-BiGRU-RNN classification model was proposed for multi species of citrus diseased leaves. The model adopted BiGRU and RNN structures in the encoding and decoding module, which can effectively extract the deep features of spectral information by using the correlation of spectral information in the front and back bands of hyperspectral images. According to the difference of spectral information of different bands, the attention mechanism was introduced to dynamically allocate weight information to improve the contribution of important spectral features to the classification model and enhance the classification accuracy of the model. Hyperspectral information of six types of citrus leaves was acquired to construct the experimental sample set, and Att-BiGRU-RNN, VGG16, SVM and XGBoost were used to establish classification models of citrus diseased leaves respectively. The overall accuracy (OA) of the Att-BiGRU-RNN model can reach 98.21% on average, which was 4.71 percentage points, 10.95 percentage points and 3.89 percentage points higher compared with that of the other three models respectively, and the recognition accuracy of herbicide and soot disease with high spectral curve coincidence was significantly improved. The experimental results showed that the deep learning method can effectively use the correlation information between different hyperspectral bands, and the classification accuracy was greatly improved compared with the machine learning method, which provided a method for rapid non-destructive detection and prevention of citrus diseases and pests.