Machine Learning and Data Processing

Deep machine learning

A core element of early detection of wildfires is our deep machine learning smoke detection algorithm. Machine learning is at the intersection of statistics, computer science and optimization. A machine learning model is able to learn representations of real world artifacts - whether it is words, images or numeric data. Traditional statistical models are expanded and optimized to develop more and more complex representations of the real world. These methods have evolved with computer science into the field of machine learning. Deep machine learning models are highly complex with millions of parameters and often require specialised computers to optimize them for production. Smoke detection is a kind of image processing task. Image processing models in deep learning have many layers.

Each layer represents different features of the image at higher and higher levels of abstraction. An archive of smoke images is used to train the parameters of our deep learning model so that it can be used to detect the presence of smoke in new images.


Real Time Intelligence

Images are ingested in real time from multiple different platforms at various altitudes, i.e. satellites, planes, high altitude drones, and ground based cameras. Visual and infrared spectra are captured in the images, which are then fed into our deep learning models.

Within minutes the images are processed for the automated detection of smoke or other features that indicate the ignition of fires.


Fire Information Acquisition and Management System (FIAMS)

Images are continuously received from scanning cameras. These images are decomposed into sub-images which are processed using the above mentioned algorithm for the presence of smoke or fire.

Additional post-processing is applied to compare recent images of the same view and other contextual information. When the smoke probability passes a threshold value and is supported by contextual information, a notification is sent to the end user of a potential fire hazard.

FIAMS collects data, creates and disseminates electronic maps in real time. This system is fully automated and connected to a network of sensors and cameras. The automated fire detection significantly reduces the response time rather than relying on current systems that rely on casual observers or fire towers. FIAMS can detect a fire as early as 3 minutes after ignition. By comparison, relying on the human eye, it currently takes 15 minutes or more. FIAMS also directly supports firefighters, commanders and agency managers. By supplying intelligence such as fire location, flame size and rates of spread FIAMS improves the development of tactical, strategic and life-saving actions.

Fireball.International has built FIAMS to provide critical information earlier in the development of the fire thus reducing the information lag for air and ground based forces managing fires. Thus leading to a significant reduction in:

  • number of lives lost
  • damage to properties and infrastructure
  • damage to ecosystems, watersheds, agriculture and
  • damage to the tourism industry