Rice disease and pest identification integrating albert pre-trained language model and improved BILSTM

To solve the low recognition accuracy and slow recognition efficiency in traditional rice disease and pest recognition technology, this study adopts a bidirectional encoder representation pre-training model from a transformer for preliminary recognition of rice diseases and pests. At the same time, a bidirectional long short-term memory network is introduced for further recognition, and the model is optimized using conditional random fields to design a fusion algorithm for rice disease and pest recognition. The outcomes denoted that when the learning rate was 0.0001, the loss of the fusion algorithm was 0.04, indicating its high accuracy. In the identification of 6 types of rice diseases and pests, the average training time of the fusion algorithm was 31.4 seconds, the central processing unit occupancy rate was 94.3%, and the memory occupancy rate was 66.4%, proving that the algorithm had high efficiency in disease and pest identification. On the PlantVillage dataset, the accuracy of the fusion algorithm was 94.3%, higher than other algorithms, indicating its good recognition performance. The fusion algorithm effectively improves the rice disease and pest identification accuracy and efficiency, providing strong technical support for preventing and controlling other agricultural diseases and pests.