Fighting wildfires with algorithms
Artificial intelligence (AI) could be the next-generation tool that helps minimize the scale of wildfire damage and reduce the cost of protection through early detection of fire signatures. This white paper surveys the current state of AI in wildfire management and makes recommendations for next steps.
While some interesting experimental AI work is being done in the area of wildfire prediction (e.g. mapping fuel moisture levels), many experts are focusing on AI as a tool to improve detection. The premise is simple: early detection gives fire crews a chance to assess the threat, send the appropriate equipment and personnel to the correct location, and contain the damage.
Generally, the AI wildfire detection process works like this. Large volumes of visual content (e.g. satellite images or live CCTV feeds) containing both positive and negative wildfire signals are collected and labeled. These data are then used to train an algorithm that locates indicators of wildfire activity based on the parameters in its database. When a potential match is found, the system generates a message with an image and location of the fire.
As a wildfire detection tool, AI is still a maturing technology. However, thanks to a trained database with more than 1.1 million images, our system operates with an industry-leading 1% false positive rate. As the database grows and the training deepens, we anticipate a steady increase in accuracy rates over the next 6 to 12 months.
Managing AI Outputs
Today, outputs from state-of-the-art AI technology offer significant fire detection value. We also see an opportunity to further improve the efficiency of wildfire management operations in camera-based monitoring stations and lay the groundwork for greater AI accuracy in the future. In this scenario, we envisage new staff priorities: AI assists the CCTV management platform rather than replacing it. Station staff spend the bulk of their time monitoring and evaluating AI alerts rather than looking for fire signatures in a 24-7 CCTV feed.
This reorganization could yield two positive outcomes. Staff are able to locate, inspect, and respond to a greater number of actual fire signatures, thereby reducing the number of fire brigade responses to false positives.
In addition, staff could contribute to the AI accuracy improvements. To decrease the number of false positives, algorithms need to be trained. That is to say, AI algorithms need more and better baseline data in order to accurately identify true fire signatures. AI training is a simple but time-intensive process. Wildfire management staff would review each AI fire warning message, evaluate its accuracy, and then enter corrected data if required. In effect, humans teach the computer how to interpret data with greater precision.
In the short term, building a better baseline can improve outputs. Over the long term, new and more robust algorithms may be required to interpret and combine a variety of data sets, including satellite images, camera feeds, and weather information. If the ultimate aim is to deploy a reliable and accurate autonomous AI wildfire detection system, it will be better served for data from multiple sources to be collected and analyzed.
We’re a technology startup from South Korea. Our domain is artificial intelligence (AI) and augmented reality (AR). Supported by a team of experts in computer vision, deep learning, graphics, data science, and software engineering, we solve problems.
Since 2016, Alchera has empowered businesses to launch products, improve efficiencies, and bring the power of visual AI to their organization.