A deep neural network approach for regional-scale 30-day accumulated urban fire occurrence forecast

Abstract

Efforts to predict and mitigate urban fires require a comprehensive approach that accounts for the complex factors driving diverse fire occurrence patterns. Previous research on urban fire occurrence prediction models has been limited, particularly in addressing both spatial and temporal dynamics effectively. This study bridges this gap by developing a deep neural network (DNN) model specifically tailored to forecast a 30-day accumulated fire occurrence map with a resolution of 2.5 km × 2.5 km for the main urban area of Hangzhou, China. Utilizing a comprehensive 9-year dataset from 2015 to 2023, we develop a novel framework that integrates heterogeneous data, including meteorological factors, urban land use, and preceding daily fire occurrence maps. The framework not only forecasts fire occurrence maps but also provides valuable insights into how different urban land use categories contribute to the spatial patterns of fire occurrences. Our results demonstrate that the model effectively predicts the total count and spatial distribution of fire incidents. Additionally, investigations into urban land use vulnerability reveal that street and transportation zones are particularly susceptible. The proposed methodology addresses the critical need for short-term fire risk forecasting, offering an effective decision-support tool for local fire departments to implement fire risk mitigation strategies.