Forest Fire Area Prediction Based On Support Vector Machine

Yunnan is one of the major forest areas in China. Its unique climate types produce many inflammable goods, thus made the forest fires numerous and varied. The study of the laws of the forest fires’ occurrence and development, and the accurate prediction of forest fires is an effective way to protect forest resources as well as people’s lives and property. In the paper,10years of the forest fires from2001to2010in Yunnan Province are analyzed and the model using the meteorological data to predict the area of forest fires is established.The paper compares the prediction effects of three forest fire models, respectively, the linear model, based on the Gaussian kernel support vector machine (SVM) model and a custom kernel SVM model. The modeling method of semi-definite programming is applied to construct the custom kernel and a grid search algorithm is used to search the optimal parameters to establish the custom kernel SVM model.The model assessment shows that the custom kernel SVM model has a smaller mean square error value than the linear model, and the Gaussian kernel SVM model, and do not produce the Learning Phenomenon. The paper uses regression error characteristic curves to compare the accuracy of three models given the certain error limit. The result shows that when the allowable error is less than0.7, Gaussian kernel and the custom kernel have the similar accuracy, and when allowable error increases, the custom kernel is much more accurate than the Gaussian kernel. The paper also designs and implements a forest fire forecasting system based on the three models, which makes the forest fire forecast convenient when using meteorological data, and provides a theoretical guidance and technical basis for using data mining methods to built forest fire prediction models.