Case study - Computer Vision for Wildfire Detection

Challenge

Wildfires are a major environmental and health risk, with a frequency that has increased dramatically in the past decade. In 2017, for example, over 9,000 wildfires destroyed nearly 6,000 square kilometres of land, killed 47 people and costing an estimated $18 billion in California alone. Early detection is one of the best ways to control the risk that wildfires pose, however the most common method for wildfire discovery and response is members of the public making reports from first-hand observations. S2DS Project Partner HAL24K posited that watching out for wildfires from a better point-of-view would help expedite early detection.

This project was aimed at creating such a solution by developing an automated detection system based on the real-time data feed generated by satellites. The small size of those images and the low accuracy of the available data posed significant challenges. The objective of the project was exploration, to gauge whether an approach to wildfire detection that is carried out from scratch, and different from the existing ones, can lead to new applications and solutions, ultimately leading to better outcomes.

Approach

The preliminary research showed that GOES satellite data from the National Oceanic and Atmospheric Administration contained ground and atmospheric temperature information and could be used to infer the type and amount of vegetation, the cloud cover and the concentration of carbon dioxide.

Thus, under the assumption that both wildfires and the specific conditions that led to them (e.g. dry vegetation) were encoded in GOES data, the team of four data scientists approached this as a classification problem, where the outcome was a system capable taking a set of images as input and classifying them into fires and non-fires.

GOES produces images every 15 min in 16 channels, covering the full western hemisphere. The team designed an API to fetch a cropped image, across any choice of time and geographical coordinates, and accounting for the satellite calibration. A scraper was then written for extracting data from Wikipedia about Californian wildfires. The team developed a framework to obtain a list of positions and times for each class, randomly sampled between 2017 and 2019.

The final piece of infrastructure combined the sampler and the data handler to produce large training and test samples, which were stored in a standard format to allow for reuse and manipulation. From this, the team produced circa 10,000 sample frames which involved 6TB of downloaded data, as well as air pollution quality data. The analysis was done using Azure and HAL24K's proprietary Dimension cloud services.

Results and Impact

The team tested a number of different approaches, including convolutional neural networks, decision trees and random forests, to achieve an accuracy of up to 85%. This was further improved by making Bayesian inferences from overlapping frames.

The approach described here is more contextual than existing detection methods and has the potential to address a wider range of applications in a comprehensive software package; for instance, the results suggest that it may be possible to identify high-risk areas inaddition to detecting already alight fires. Additionally, one of the models developed could be used for damage estimation after the event.

Data-driven, intelligent, wildfire solutions will protect the lives, property, and livelihoods of the generations to come through better prediction and detection of fire. This will also further improve protection by allowing insurers to expand their coverage and optimise their services. The project is especially relevant given changes in Earth’s climate, and the increasing risk of extreme temperatures and weather.

“This is the fourth year that HAL24K has sponsored an S2DS project and we always aim to challenge our team to deliver innovative solutions. The team were tasked to use high-cadence satellite data, in conjunction with other data, to prototype a way to detect wildfires. The team delivered; not only did they build a proof-of-concept satellite wildfire detector but they did so in a methodical way that delivered an initial codebase including documentation. HAL24K will now build upon these initial products and incorporate the components into our Dimension platform to enable clients to build their own advanced data science solutions.”

Dr Adam Hill, Lead Data Scientist, HAL24K