基于自适应判别深度置信网络的棉花病虫害预测

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

作物病虫害预测是病虫害防治的前提,利用深度学习预测作物病虫害是一个有效且具有挑战性的研究课题。该文针对深度置信网络(deep belief network,DBN)在作物病虫害预测中的训练耗时长和容易收敛于局部最优解等问题,将自适应DBN和判别限制玻尔兹曼机(restricted boltzmann machine,RBM)相结合,利用棉花生长的环境信息,提出一种基于自适应判别DBN的棉花病虫害预测模型。该模型由3层RBM网络和一个判别RBM(discriminative restricted boltzmann machine,DRBM)网络组成,通过3层RBM网络将棉花生长的环境信息数据转换到与病虫害发生相关的特征空间,通过自动学习得到层次化的特征表示,再由DRBM预测棉花病虫害的发生概率。该模型将自适应学习率引入到对比差度算法中,通过自动调整学习步长,解决了在传统DBN模型训练时学习率选择难的问题;在学习过程中通过在DRBM中引入样本的类别信息,使得训练具有类别针对性,弱化传统RBM无监督训练时易出现特征同质化问题,提高了模型的预测准确率。对实际棉花的"棉铃虫、棉蚜虫、红蜘蛛"虫害和"黄萎病、枯萎病"病害的平均预测准确率为82.840%,与传统BP神经网络模型(BPNN)、强模糊支持向量机模型(SFSVM)和RBF神经网络模型(RBFNN)分别提高19.248%,24.916%和27.774%。

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.