Content overview : On August 8, local time, a wildfire broke out in Hawaii, USA, and local residents and tourists had to jump into the Pacific Ocean to avoid the fire. As of August 17, the wildfire has killed 110 people and more than 1,000 people are missing. At the same time, the United States, Canada, France and other places are also being hit by wildfires. Wildfires are relentless, and it is difficult for people to respond quickly in the face of sudden wildfires. Now, with the help of AI, there are new advances in the detection and prevention of wildfires.
Keywords : wildfire artificial intelligence explainability AI
Author｜
Editor｜Sanyang
This article was first published on the HyperAI Super Neural WeChat public platform .
The earth is irreversibly experiencing global warming. According to statistics from the Copernicus Climate Change Service (C3S) of the European Union, July 2023 will be the month with the highest global average temperature since 1940. The temperature has risen by about 1.5°C compared with the average temperature during the pre-industrial revolution, surpassing the Paris Agreement. 》The threshold value set.
Figure 1: Global average temperature from 1940 to July 2023 [1]
The most intuitive feeling brought about by global warming is high temperature. In densely forested areas, hot and dry climates can easily trigger wildfires (Wild fire) . On August 8, a wildfire broke out in Maui, Hawaii, USA. The wildfires took advantage of Typhoon "Dora", quickly swept through the forest, and spread to the local cultural center - Lahaina, leaving tens of thousands of people homeless.
Meanwhile, wildfires are raging in the dry western part of North America. The fire map of the Fire Information for Resouce Management System US/Canada (FIRMS, Fire Information for Resouce Management System US/Canada) shows that in the past week, more than 1,000 acres of wildfires have spread in western Canada, and there are also a large number of fire points in the eastern United States .
Figure 2: Map of fires in North America in the past week[2]
The wildfires are fast and relentless. In the face of sudden wildfires, it is difficult for people to respond in a timely manner. But now, we can use AI to monitor and predict wildfires in real time to minimize the losses caused by wildfires .
Forecasting the Wildfire Trident
Geographic data: Australian universities develop XAI
In May 2023, Abolfazl Abdollahi of the Australian National University and Biswajeet Pradhan of the University of Technology, Sydney, used interpretable AI (XAI) to conduct a comprehensive analysis of different geographical factors in Gippsland, Australia, and obtained the local wildfire probability distribution , providing a new way to predict the occurrence of wildfires.
Figure 3: XAI Workflow
for Predicting Wildfire Probability Geographic features that have a significant impact on wildfire probability include environmental, geological, vegetation, and meteorological factors . In this study, the researchers mainly used the following 11 features, including precipitation, wind speed, air temperature, humidity, vegetation distribution, vegetation area, fuel nitrogen in plants, moisture, and the altitude, slope, and orientation of the area.
The study selected 521 wildfire occurrence points in Gippsland City, and trained them on the Moderate Resolution Imaging Spectrometer (MODIS) data, thermal anomaly data and fire history data from 2019 to 2020, and through cross-validation and retention data sets, the The training set is processed to ensure the accuracy of the training process.
First, the cross-validation technique randomly divides the training set into 5 subsets, of which 4 subsets are used for model training and 1 subset is used as a validation set.
Figure 4: Model training process. The blue data is used for training, and the orange data is used for verification.
After the training is completed, the environment, geology, vegetation, and meteorological characteristics of Gippsland are input into the model to obtain a complete probability map of wildfire occurrence in this area. As shown in the figure, the model The prediction results are basically consistent with the historical wildfire areas in this area, indicating that XAI can effectively predict the occurrence of wildfires based on geographical features .
Figure 5: Comparison of model prediction results and historical wildfire areas
a : The wildfire occurrence probability map predicted by the XAI model;
b : Map of the historical wildfire area of Gippsland.
Paper address:
https://www.sciencedirect.com/science/article/pii/S0048969723016224
Video Data: Identifying Fire Points Based on Smoke
Wildfire forecasts based on geographic data can only increase people's vigilance, but cannot provide real-time observation of wildfire occurrence. To this end, the California Department of Forestry and Fire Protection, in cooperation with the University of California, San Diego (UCSD), developed a wildfire prevention project called ALERTCalifornia based on AI. This project monitors the local area through more than 1,000 cameras distributed in California, uses AI to identify abnormal conditions , sends an alarm to the emergency command center, and reminds the on-duty personnel to confirm whether there may be wildfires.
Figure 6: ALERTCalifornia camera distribution and real-time images
This project was officially put into operation in July 2023, and it came in handy soon. At 3 a.m., a camera detected the fire in the Cleveland National Forest, 80 kilometers east of San Diego. Since the incident occurred late at night, the smoke was difficult to detect, and it was difficult to identify the ignition point with the naked eye, which could easily lead to the spread of the fire. But the AI alerted the fire chief in time and helped the fire department put out the wildfire within 45 minutes.
However, this technology also faces many challenges in the development process. One of them is how to make AI accurately distinguish wildfires and other disturbance factors and make accurate judgments . There are plenty of things in the forest that can trigger false alarms, including wildly shaped clouds, dust in the air, and even exhaust from passing trucks. With the joint efforts of more than a hundred experts, after several weeks of training and iteration, the accuracy of AI has been significantly improved.
ALERTCalifornia address :
https://cameras.alertcalifornia.org/
Satellite data: Near real-time wildfire monitoring with secondary screening
In specific areas, cameras can effectively monitor wildfires, but this method is difficult to promote on a large scale, especially in some areas with vast areas and complex terrain, the cost of deploying and maintaining cameras will increase significantly. Therefore, real-time satellite data can also be used as an AI weapon for real-time monitoring of wildfires .
He Binbin's research group at the University of Electronic Science and Technology of China built a machine learning model combining Random Forest and Spatial Contextual Algorithm , and realized near real-time wildfire monitoring through satellite data.
Traditional AI wildfire identification systems often use a single algorithm, resulting in data omission or false positives. In this study, the data were first strictly screened through the random forest model to avoid omissions. Then use a spatial context algorithm with a relatively low threshold for secondary screening to eliminate false alarms , thereby improving the accuracy of the monitoring model.
Figure 7: Workflow of the wildfire recognition model
The researchers trained the model on wildfire data recorded by the Japan Aerospace Exploration Agency (JAXA) and NASA. Satellite data of wildfire locations from 2020 to 2022 were then used as validation data for the model to make a judgment call.
Figure 8: Satellite data verification results of the wildfire identification model As
can be seen in the figure, for most fires, this model can accurately locate the source of wildfires and make timely early warnings . The AI wildfire recognition model uses real-time satellite data as the data source, combined with the random forest model and spatial context algorithm, which reduces the false negative rate and false positive rate of the model, and realizes near real-time wildfire early warning.
Paper address :
https://www.mdpi.com/2272228
Wildfire Fighting: A Race Against Time
The primary cause of wildfires is human activity , such as unattended campfires, littered cigarette butts, or improper use of equipment. The likely cause of the Hawaiian wildfires was damage to electrical facilities. Under the joint influence of the high temperature and dry environment brought about by global warming and typhoon "Dora", the flames developed rapidly, causing huge losses to the local residents.
Figure 9: Lahaina after the wildfire [3]
Fighting wildfires is a race against time . It is estimated that wildfires can spread about 23 kilometers per hour, and the area burned will quadruple in size within 4 hours. In 2017, the Thomas wildfire in California, USA, could spread an area of a football field per second, and eventually continued to burn for more than 3 months.
Therefore, the monitoring and fighting of wildfire often requires the use of multiple technologies to find the fire point in the shortest possible time, extinguish the wildfire as soon as possible, and avoid its spread. At present , China has six lines of defense for monitoring wildfires, namely satellite monitoring, aircraft monitoring and patrolling, observation towers in forest areas, video surveillance systems, ground patrolling and network public sentiment awareness .
With the support of various technologies, the number of forest fires in my country has dropped from 7,723 in 2010 to 709 in 2022. However, the forest fire in Muli County, Sichuan Province in 2019 and the forest fire in Xichang City, Sichuan Province in 2020 both caused heavy casualties, and the monitoring and prevention of forest fires still face challenges.
At present, geographic data, video data, and satellite data can all be used as the raw data of AI for early judgment of wildfires to kill the fire in its infancy. Although my country has a vast territory, the climate and topography of different regions are quite different, so it is difficult to completely prevent wildfires.
This article was first published on the HyperAI Super Neural WeChat public platform .
Reference link:
[1]https://climate.copernicus.eu/july-2023-warmest-month-earths-recent-history
[2]https://firms2.modaps.eosdis.nasa.gov/
[3]https://www.washingtonpost.com/climate-environment/2023/08/10/hawaii-wildfire-maps/
[4]https://www.bjnews.com.cn/detail/168429261314778.html