In recent years,the number of occurrences of forest fire has been increased,which causes serious damage to the ecosystem,indirectly causing global warming and extreme weather. There is a clear feedback between climate change and forest fire. Climate change induces weather extremes,which leads to forest fires,while fire emissions contribute to climate change. Accurate prediction of the forest fire can enable relevant personnel to take effective prevention and control measures in advance. Controlling the occurrence of forest fire not only protects the ecological environment, but also greatly addresses climate issues,so accurate prediction of forest fires is of great significance. Traditional forest fire prediction models are mostly mathematical methods and shallow neural networks. They take the meteorological factors as the inputs to predict forest fire. When the amount of data increases,they are prone to be problems, such as modeling difficulties and reduced prediction accuracy. Deep learning has certain advantages in dealing with a large amount of nonlinear data,and its model has a multi-layer network structure. By training the deep learning model, more representative feature values can be extracted and the implicit relationship among data can be found,to achieve the purpose of accurate classification and prediction. Compared with the traditional fire prediction model,the deep learning model has a deep network,and the parameters can be adjusted autonomously through the training data, which is suitable for a large amount of nonlinear data. Therefore,this study used the deep belief network (DBN) in the deep learning model as the predictive model and the meteorological factor as the input data to solve the problem that the traditional forest fire prediction model could not predict well when faced with large amounts of data.In this study,the SMOTE (synthetic minority oversampling technique) algorithm was combined with DBN to balance the forest fire data set and increase the amount of training data,which made the forest fire data more suitable for DBN. As a result,the prediction accuracy of forest fire was improved. The prediction result of the model was compared with the support vector machine (SVM) and back propagation (BP) models. The modeling results showed that the prediction accuracy of the model can reach 84%,which was obviously better than SVM and BP network. It was demonstrated the superiority of applying deep learning to forest fire prediction. This study can provide a theoretical foundation for the application of deep learning in the field of forestry.