Part 2 - What can AI see?
Case Study: An Early Warning Forest Fire Detection System
The problem
As global warming intensifies, and extreme heat events become more prevalent, the frequency and severity of forest fires have increased. Over the past 20 years, forest fires have accounted for more than a quarter of tree cover loss; 9.3 million hectares in 2021 alone (MacCarthy et al., 2023). These incidents cause serious damage to our ecosystem and lead to a significant loss of life in the communities that live within and near forests. Additionally, forest departments lack the resource to continually monitor the entire forest and identify problems early enough to contain them.
The idea
In 2021, the Lahore University of Management Sciences (LUMS) partnered with WWF, FCDO, and the Frontier Tech Hub to explore how computer vision could help mitigate this challenge. Over the past 2 years, the team have been developing an early warning system which uses data gathered from IoT sensors and cameras to:
Predict where fires are likely to arise
Identify the breakout of fires early
Track their spread to mitigate their damage
Early warning forest fire systems system are not a new idea. However, the team recognised the need to develop a system tailored to the local context, economically viable, and useful for forest fire staff on the ground.
How does the tech work?
As mentioned earlier, complex use cases such as a system capable of detecting fires, involve multiple computer vision tasks. In this case, LUMS developed the system with two underlying techniques:
Binary classification: A binary classification algorithm was used to identify whether an image was either: (a) An image showing fire/smoke, or (b) an image not showing fire/smoke. This classification allowed them to flag specific images for review by the forest department guardians on the ground.
Object detection: They combined this with an object detection layer which was used to identify the specific parts of the image which included the fire or smoke. By identifying where in the image the fire had been spotted, they could zoom in and take clearer images of the potential fire for further analysis.
The system first identifies whether an image potentially contains a fire using the classification layer, and then uses object detection to determine the location of that fire within the image, such that the forest fire operators know where to direct their response.
In the next part of this case study, we’ll explore some of the ways that the team built this system. Click below to continue.