Construction of forest fire prediction model based on Bayesian model averaging method: taking Dali Prefecture, Yunnan Province of southwestern China as an example

  Objective  Based on the Bayesian model averaging method and binomial Logistic regression model, this paper constructs a forest fire prediction model in Dali Prefecture, Yunnan Province of southwestern China, so as to improve the prediction accuracy of forest fire and provide technical support for forest fire management in the study area.

  Method  Using the forest fire data and corresponding meteorological data of Dali Prefecture from 2000 to 2013, the binomial Logistic regression model and the Bayesian model averaging method were used to empirically analyze the response of forest fires to meteorological factors in this area. The binomial Logistic regression model is a single model. Before modeling, the explanatory variables with significant collinearity were eliminated by multicollinearity test. Then, the final variables were screened by stepwise regression method and the parameters were fitted. The Bayesian average model is a combined model. When modeling based on the Bayesian model averaging method, the Occam’s window method was used to appropriately adjust the model space, and the posterior probabilities of the five optimal models were used as weights for weighted modeling. In this paper, the all sample data were randomly divided into 80% training samples and 20% test samples. A model was built based on the training samples to predict the test samples. The accuracy of the model was calculated by comparing the observations and predictions.

  Result  Fitting through the binomial Logistic model, the results showed that: the model fitting goodness was 0.783, and the prediction accuracy was 0.718; through the Bayesian average model fitting, the results showed that: the model fitting goodness was 0.868, and the prediction accuracy was 0.807. The comparison of the prediction results of the two models showed that: in the training set, the prediction accuracy of the Bayesian average model was 9.3% higher than that of the binomial Logistic regression model; and in the test set, the former was 8.9% higher than the latter.

  Conclusion  In the prediction model of forest fire occurrence in Dali Prefecture based on meteorological factors, the goodness of fit and prediction accuracy of Bayesian average model were higher than that of binomial Logistic model, indicating that the Bayesian model averaging method had certain practical application significance. It can be used to improve the prediction accuracy of forest fire in the study area, which is beneficial to the decision management of forest fire.