Data mining techniques in the assessment of usability and effectiveness of forest fire video surveillance

Abstract

Since forest fires are natural phenomenon that cannot be avoided nor their occurrence can be prevented it is of utmost importance to have a reliable fire management system at hand. Damage prevention and early fire detection are two approaches that are widely used in forest fire management today. In the last decade, video surveillance has been recognized as useful tool for forest fire management. It helps both firefighters and forest managers in several aspects of their work, and consists of a video sensor located in the area that would otherwise be difficult to access. This sensor provides video information about the state of the surveilled forest, and can be helpful in two forest fire video surveillance phases: • monitoring phase, that takes place before fire occurs band includes continuous monitoring of the surrounding area in search for signs of fire and smoke, and • surveillance phase that occurs during the firefighting operative actions and that provides commanding officers with information on the severity and spread of fire. Since there is a great demand for video surveillance for early fire detection, both in the firefighting and in the public community, this area of research is in the focus of many scientists and developers. Automatic fire detection can aid early fire response, therefore minimizing the possible fire damages while also not relying on human observers that might be suspect to inadequate experience and possible tiredness. The work presented in this paper was motivated by the need to define quantitative measures that can help decide if a video surveillance system is effective in firefighting processes during the fire season. These measures can ultimately be used as a tool to perform cost-benefit analysis when designing and planning an integration of firefighting and forest fire management with systems for video monitoring and video surveillance. A pilot forest fire video monitoring and surveillance system was installed in Šibenik-Knin County as the output of the HOLISTIC project at the end of 2016 fire season. In 2017 fire season the system was fully functional and used in everyday activities. In this paper we used a variety of available data related to forest fires that took place in the area of Šibenik-Knin County in the summer of 2017. All available data was aggregated and used as the input data for the presented analysis. Several data mining techniques were used for data formatting and association. Ultimately this data was combined in a single document where the activities that were logged in the system logs were correlated with information from the operative actions reports, social media and other sources. Data mining techniques such as clustering and frequent itemset analysis were utilized for the assessment of usability of the overall video surveillance system. The correlation of the alarms raised by the system and real operative actions in the context of media reports provide a measure of usefulness of the video surveillance system both for fire detection and for guidance during the firefighting process.