Urban Fire Situation Forecasting: Deep sequence learning with spatio-temporal dynamics

Abstract Understanding the evolving discipline of urban fire situations is a basic but challenging task for urban security and fire-fighting decisions. Traditional methods forecast the urban fire situation through mathematical modeling and statistical learning, which could be interpretable but generally lack of efficiency and practicality. Recently, some deep neural network methodologies, especially convolutional neural network (CNN) and recurrent neural network (RNN), are presented as paradigms to capture dynamics in spatial–temporal complex phenomenon, which tally with the characteristics of fire situation forecasting. In this paper, we propose a novel deep sequence learning model as the fire situation forecasting network (FSFN) to better process the information and spatio-temporal correlations in regional urban fire alarm dataset. FSFN model integrates structures of Variational auto-encoders and context-based sequence generative model Seq2seq to obtain the latent representation of the fire situation and learn the spatio-temporal dynamics. Furthermore, we augment the network structure of FSFN from a simple deep sequence generative model to adversarial fire situation forecasting network with auxiliary information(Adversarial FSFN-A). The experimental studies demonstrate the effectiveness of Adversarial FSFN-A has superior spatio-temporal distribution prediction of multi-type urban fire situation.