Early detection of wildfires is vital to reduce fire caused deaths and property loss, and to prevent disastrous impact on wildlife and the environment, as well as to minimize the economic loss of the government on firefighting. Traditional approaches to wildfire detection often have delays, for example, fire observation towers take time to install and run but they only cover a small area; others may tend to be uneconomical, such as spotter aircrafts which are expensive to run but only operable for a short time.
In recent years, more satellite data on earth surface observation become available and this opens the door for new methods to be developed, which are expected to be prompt in time when fires are still small. However, the current satellite image-based detection methods are still ineffective for early fire detection due to low spatial or temporal resolution of the sensors. That says, wildfires usually cannot be detected directly by the satellites until they have burnt for a relatively long time and to a relatively large scale.
Instead of direct fire detection, smoke detection is expected to an alternative method since smoke disperses very fast into the air and can be visibly detected by the satellites quicker than fire. However, there remains many challenges. First, smoke has similar characteristics with cloud, dust, haze, which are difficult to be visually distinguished most of the time. Due to this, the false positive rate of smoke detection remains high for most developed models. Second, smoke does not have fixed shapes, which usually change very fast especially when there are winds. This will dramatically increase the complexity of the smoke detection models. Furthermore, the characteristics of smoke are closely associated with the fuels and are subject to the local environmental conditions, which makes the detection models hard to be applied to a different area without further adjustment. The differences between the sensor specifications also raise challenges for model reapplication and integration.
This project aims to develop practical machine learning technologies which can address the above challenges for smoke detection based on satellite imagery. The technologies are expected to become a new way to address the fire detection problems in vast remote areas. The application of the technologies is hoped to reduce the cost of running remote fire towers, to mitigate the risks on people working at the towers, and to shorten the decision time between the start of the fire and proper reactions taken.
Associate Professor Jixue Liu, University of South Australia
Liang Zhai, University of South Australia