Fire spread prediction model based on multi-scale convolutional neural network

Forest fire spread prediction is a very tricky problem, and commonly used models are difficult to apply to real forest fire cases due to their low prediction accuracy. This paper presents a method for estimating the time-resolved spatial evolution of forest fires using an optimized multi-scale convolutional neural network(OMS-CNN), whose hyperparameters are processed using the particle swarm algorithm. The multi-scale convolutional structure can extract both the global and local characteristics of the fire scene. The particle swarm algorithm can select the appropriate hyperparameters to maximize the model performance and minimize the loss value. The OMS-CNN model was trained and tested in order to cope with the spread of forest fires under field conditions. Outdoor burning experiments were designed and dataset used for training, validation and testing was recorded by a UAV equipped with an infrared camera. The comparison of the OMS-CNN model with the two degenerate models and three advanced wildfire models in terms of Kappa and Hausdorff distance values demonstrates that the OMS-CNN model achieves higher predictive accuracy. Furthermore, the input variable ablation experiments confirm the model's practical value in real-world scenarios. The model can be applied in forest fire managing and fighting system, which has extremely important theoretical significance and practical application value.