Abstract: Abstract: In order to quantitatively analyze the relationship between the regional economic development level, population density, building density, fire station distribution and other factors and fire occurrence, a variety of machine learning classification algorithms are introduced for research. Use ArcGIS 10.2 toolbox to process non-numerical data, and classify according to the level of fire core density in fishing nets, so that variables are converted into corresponding numerical data; under the condition of ensuring accuracy, use multiple random forest algorithm to perform feature screening, and perform deep learning training on the remaining features after screening. At the same time, the support vector machine algorithm is used to train all features, and the prediction models are constructed respectively. Finally the three algorithms are fused by weighted average. The ROC curve and the classification accuracy of the 4 models are compared. Taking the real fire data collected in the Chongqing fire alarm system as an example, the results show that the accuracy of the four models are all higher than 90%; the accuracy and Kappa values of the models after the coupling of the three algorithms are 0.980 7 and 0.843 6, and the result is more stable and accurate.

Key words: Key words: regional fire risk, support vector machine, random forest, deep learning, machine learning, ArcGIS