Abstract:  The occurrence of flying fire in forest fire greatly accelerates the spreading speed of forest fire and seriously threatens the life safety of firemen. It is of great significance to forecast and early warning of spot fire. Most studies ignore the correlation between physical characteristics in the formation of spot fire, it is not enough to provide effective judgment basis for its prediction. A prediction model of spot fire based on linear-chain conditional random field is proposed. Combined with the temporal characteristics of the formation of spot fire, the prediction of the state sequence of each stage in the formation of spot fire is completed, and the prediction of spot fire is realized. The experimental results show that the model has high prediction accuracy, which is basically above 75%. The response time is controlled within 10 ms, which is lower than the time difference between two spot fire and the formation of fire lines. To a certain extent, it can meet the real-time requirement. It provides useful key information for the efficient fighting of natural fires in forest and grassland, and also provides important decision-making basis for the command and dispatch of fire fighting.

Key words: linear chain, conditional random field, spot fire, batch size, learning rate