Residential Fire Analysis and Probability Mapping
Sikkerhed og sundhed
While fire prevalence in Denmark is on the decline, in the period of 2010-2018 the country still recorded over 58,000 building fires, including nearly 600 fatal fires. With the combination of predictive analysis and detailed geodata, there are more possibilities than ever to uncover the primary causes of residential fire in order to prioritize resources for preventive measures in areas with high fire vulnerability.
The Danish Emergency Management Agency and the University of Copenhagen have developed a project which incorporates fire, socio-demographic and building / property data into statistical and machine learning analysis to both uncover factors that contribute to fire prevalence as well as to create a mapping tool that can assist fire brigades in visualizing areas of high residential fire risk. These areas can be pinpointed for smoke alarm distribution and fire prevention education.
Målgruppen for indlægget er deltagere ved Kortdage 2021 fra den offentlig administration i kommuner, regioner og stat.
Yderligere uddybning af abstract
Statistics show that residential buildings contributed to the highest proportion of Danish building fires between the years of 2010-2018. This, in combination with published knowledge on the link between human factors and fire risk, implies that investigating the variations of socio-demographics in residential areas may reveal insight into where more fire safety resources are needed.
This study makes use of a dataset consisting of historical fire incident records from the Danish Emergency Management Agency, socio-demographic variables from the Nordic data and analytics company, Geomatic, and building / property data from the Danish BBR and ESR. The dataset spans the whole of Denmark’s populated areas on a 100x100 meter grid, allowing a detailed analysis of variable influence in the study area.
There are two phases of analysis in this project: One phase focuses on creating an explanatory model using linear regression techniques. The primary focus of this phase is to identify the strength and nature of risk determinants’ relationship to fire probability. While linear regression analysis allows for a direct interpretation of the influence of the explanatory variables, non-linear machine learning algorithms often have greater success in predicting with a higher accuracy. For this reason, a variety of machine learning techniques will be used for the second phase of fire prediction mapping and visualizing vulnerable areas for municipal fire brigade usage. The outputs of both phases will thus consist of detailed fire prediction maps and information on significant variable relationships in the prediction of building fires.
Additional authors to the scientific work behind the abstract are Stefan Oehmcke, Hans Skov-Petersen, Patrik K. Nyed, Rasmus Jenle, Vibeke Ø. Thomsen, and Christian Igel.