Forecasting of cotton diseases and pests based on adaptive discriminant deep belief network

Abstract: Cotton diseases and pests seriously affect cotton quantity. Timely and accurate prediction of diseases and pests is very important for crop growers to effectively prevent and monitor cotton diseases and pests. Cotton diseases and pests can be forecast by environmental and weather information. Through various sensors in the internet of things, it is easy to acquire a lot of environmental and weather information, and many cotton existing prediction methods, techniques and systems have been proposed. However, the occurrence and development of cotton diseases and pests involve various factors, among which there are complex interactions and mutual influences. The traditional prediction model of cotton diseases and insect pests has limited expression ability and generalization ability, and the accuracy of prediction is not high. Many existing prediction models cannot meet the actual needs of pest and disease prediction system. Therefore, the prediction of cotton diseases and pests is still a challenging problem in computer vision. In recent years, deep learning networks have won numerous contests in pattern recognition and machine learning. Deep belief network (DBN) is one of the most widely used deep learning models and has been successfully applied in many fields. DBN is a superposition model composed of several restricted Boltzmann machines (RBM). However, in DBN, there are a lot of problems, such as time-consuming to pre-train, easy to get into the local optimal solution, unsupervised training and poor generalization. An adoptive discriminant deep belief network (ADDBN) is proposed to solve the time-consuming problem in the pre-training process of DBN, and then a forecasting model of cotton diseases and pests is proposed based on environmental information and ADDBN. ADDBN is constructed by three RBMs (restricted Boltzmann machine) and a discriminant RBM (DRBM). In DRBM, the label information is introduced to training process of RBM, and the discriminant information is added into learning process through constraint on the similarity of feature vectors to improve the forecasting rate. In ADDBN, an adaptive learning rate is introduced into the contrastive divergence algorithm to accelerate the model convergence. Comparing with DBN, the proposed model has two advantages, (1) adaptive learning rate is introduced into the contrast algorithm to automatically adjust learning step, which can solve the problem to choose the learning rate in the training traditional DBN model; (2) the class information of samples is introduced into DRBM in the learning process. Then the model can be targeted trained, which can weaken the characteristic homogeneous in unsupervised training the traditional RBM and improve the forecasting accuracy of the model. Finally, a series of experiments were carried out on a dataset of cotton diseases and pests to test the performance of ADDBN. The results showed that the convergence rate is accelerated significantly and the forecasting accuracy is improved as well. The experiment results on the environmental information database of "three worms and two diseases" of cotton in recent 6 years showed that the proposed prediction model has better prediction effect than the traditional prediction model such as BPNN, SFSVM and RBFNN, the prediction performance is improved by 19.248%, 24.916% and 27.774% respectively. It is an effective method to predict crop pests and diseases with faster convergence rate, good generalization ability and higher prediction effect.