Abstract: As for the large capacity and diversity of the environmental information of cotton growth related to the occurrence of diseases and insect pests, a prediction method of cotton diseases and pests was proposed by combining the environmental information and improved depth belief network (MDBN). MDBN consisted of a three level restricted Boltzmann machine (RBM) network and a BP neural network (BPNN). In the method, MDBN was used to extract the characteristic variables reflecting the occurrence of pests and diseases, and BPNN was utilized to predict the cotton pests and diseases. The characteristic of MDBN was that the adaptive learning rate was introduced into the unsupervised training stage of MDBN, and the performance of MDBN was optimized by the selection of training data batches, the iteration cycle of parameter tuning and online learning and training, so that MDBN could fully extract the information features from the dataset to improve the network precondition rate. The prediction results on the actual environmental information dataset of the cotton diseases and insects showed that the proposed method was effective for predicting crop pests and diseases.

Key words: disease and pest forecasting, cotton environmental information, restricted Boltzmann machine (RBM), modified deep belief network (MDBN)