A litany of wildfires is wreaking havoc around the world. As the frequency and intensity of wildfires rise, it is speculated that our ecosystems is perhaps witnessing a transition from regular wildfire seasons to wildfire years! The last few years in this regard have seen a significant increase in the frequency of wildfires. In April 2020, for example, fire alerts around the world increased by 13% compared to the previous year, which had already witnessed a new record for wildfires.
Learn more about this in our blog “ Decoding Wildfires: Understanding The New Normal.”
While many factors are responsible for causing blazes, climate change is known to be one of its most prominent drivers. This is because wildfires and climate change are closely associated and tend to reinforce each other. Dry and hot weather, a result of climate change, drives flammable conditions, often leading to wildfires. These wildfires further worsen the climate change crisis by emitting a massive amount of greenhouse gases. In 2021, wildfires emitted 1.76 billion tonnes of carbon globally.
The Need for Predicting Wildfires
Wildfires have a devastating impact on almost all aspects of life, ranging from their effect on public health to wildlife, from having cultural implications to impacting businesses. The economic impact of it has been dire too, with the average annual global cost of wildfires estimated to be about USD 50 billion. Read more about how wildfires are impacting businesses globally in our blog.
As the fire seasons get more extreme, it is estimated that they burn 3.5 to 4.5 million square kilometers (km2) of vegetation worldwide every year. This is equivalent to an area about the size of the entire European Union!
The crisis is expected to further exacerbate, with a study estimating that wildfires will likely increase by a third by 2050 and by 50% by the end of this century. Thus, as the crisis is poised to worsen, the current approach of fire management, which focuses on quick detection and suppression of fires, needs to be more robust. In addition, these traditional methods are short-term solutions that are proving to be uneconomical in the long run. For example, a central estimate in the United States is that the total cost of fire damages is 20 times more than the suppression costs.
Accurate fire prediction tools in this regard can create a paradigm shift in the current fire management approach and significantly minimize risks and losses. Apart from its value in holding the potential to save lives and minimize the damages caused to the ecosystem, fire predictions can especially benefit businesses significantly. With forewarning, authorities may be able to prevent fires, helping bring down the material and financial losses caused by fire damage to a company’s infrastructure and assets. Even if prevention is not possible, authorities can ramp up disaster preparedness and mitigation efforts, thus helping drastically minimize damages. It can also bring down the losses resulting from power shut-offs, business closures, and insurance payouts. The benefits of fire prediction would be significantly felt in industries that are specifically susceptible to wildfire risks, such as the insurance, banking, energy and utility, and tourism sectors.
Methodologies Used to Monitor Wildfires
Systems for predicting fire hazards are crucial for predicting fire risks as they support a fire's monitoring and extinction phase. They also help in resource allocation and planning fire management strategies. Currently, various methods are being used, including physics-based approaches, statistical models, and deep neural network models.
The links between environmental conditions, climate change, and anthropogenic factors causing wildfires and biomass burning are the subject of numerous research that is now being published. Researchers and organizations use multiple methodologies to monitor and predict probable fires. Some of the widely explored methodologies are summarized below:
1. Forest Fires Risk Probability Maps Using Fire Weather Index
The Fire Weather Index (FWI) is the most well-known model for determining fire risks. The system indicates fire danger conditions. Four weather parameters influence it:
the humidity of the air at the beginning of the afternoon (when it has its lowest value);
the temperature in the middle of the afternoon (when it has its highest value);
the 24-hour total precipitation (from noon to noon);
the maximum speed of the average wind.
FWI has been extensively researched to monitor fires based on historical fire data and weather parameters. The higher the FWI is, the more favorable the meteorological conditions to trigger a wildfire are. In a changing climate, FWI can aid in forming long-term tourism strategies and help in planning upcoming investments.
FWI, however, has two major shortcomings -
As FWI was developed in Canada, it was specifically defined and calibrated to monitor fires in the location. Its inability to provide a region-independent definition and index limits its uniform applicability worldwide. For instance, the FWI index is a whole number that ranges between 0 and 20 in France and above 30 in Canada.
FWI’s calculation relies heavily on weather forcings, and no information on the actual vegetation status is considered. FWI only considers surface weather conditions, despite the important role that atmospheric instability plays in developing very large wildfires.
2. Fire Spread Prediction Modelling Using Burnt Area
To reduce the immediate and long-term effects of wildfires, it is essential to predict the spread of the fire and the area burned. Machine learning models have been used to predict burnt areas and/or fire spread. In the paper “Next Day Wildfire Spread: A Machine Learning Data Set to Predict Wildfire Spreading from Remote-Sensing Data”, probable fires that will take place during the next day have been predicted.
The following parameters were used to train a deep-learning model:
Remote sensing data on historical wildfires,
the population density
Convolutional auto-encoder, a specialized type of neural network (a neural network is a collection of algorithms that aim to identify underlying links in a set of data using a method that imitates how the human brain functions), was leveraged to take advantage of historical wildfires with 11 observational variables laid over 2D regions at 1 km resolution and compare its performance with two other machine learning models, namely logistic regression and random forest. The deep learning model took advantage of the spatial information in the input data for prediction.
Next Day Wildfire Spread paper demonstrates the potential of deep learning approaches to predict wildfires from remote-sensing data and also illustrates the performance gaps on the example task of day-ahead fire spread prediction. An accuracy of 28.4% was seen in neural networks, and other ML models had a relatively lesser accuracy ranging between 22% and 19%.
A significant limitation in models that rely on burnt area products is that they often lack timely data availability - burnt area products become available months after a disaster, causing a delay in forecasting the effects of these products in the future.
3. Determining the Likelihood of Fires Using Fuel Load Distribution
The likelihood of a wildfire occurring depends on three factors:
the accessibility of ignition sources,
the propensity of vegetation and litter fuel to catch fire, and
the ease with which the fuel spreads once it has done so.
Live fuel moisture content (LFMC), defined as the mass of vegetation water per unit of dry biomass, is a key determinant of all three components - wildfire ignition, fuel availability, and fire spread. Previous studies of large wildfires have observed clear thresholds between fire size and LFMC. In the paper, SAR-enhanced mapping of live fuel moisture content, a physics-assisted empirical method, has been used to produce high-resolution (250 m) LFMC maps across the western US every 15 days. To estimate LFMC, Sentinel-1 synthetic aperture radar (SAR) backscatter and Landsat-8 surface reflectance have been combined with static variables in a recurrent neural network (RNN) deep learning model. The DL model had an R2 score of 0.49 and an RMSE of 25%.
Challenges and Limitations in Existing Approaches
Most currently used forest fire prediction systems rely on manually created features to make forecasts. These systems often experience several shortcomings, such as:
Fire risk prediction depends on the complex relationship between existing vegetation, fuel load, and numerous weather and soil-related parameters.
Using a one-shot set of mathematical equations to represent all the relationships becomes extremely challenging.
Fire risk models which rely on lab experiment results and static factors often suffer from erroneous predictions even if the prediction algorithm is optimal.
Many fire risk models predict risk indexes which are often not standardized across different regions and thus not comparable to other similar indices.
Indices are also harder to validate since it needs to be converted to equivalent fire counts using some conversion to compare with actual fire events per region.
How does Blue Sky Analytics do it?
At Blue Sky Analytics, we developed a Fire Predictions dataset that can forecast biomass fire events seven days into the future on a weekly rolling basis.
We combine large-scale earth observation datasets with machine learning proprietary algorithms to predict weekly fire counts on a 25 km x 25 km grid. The predictions indicate the location and number of fire events likely to occur over the next seven days. For each grid of 25 km x 25 km, the model considers various factors, including historical fire data, weather conditions, vegetation characteristics, and land use, to predict the fire count. When the data is available, the performance of each predicted fire count is evaluated by comparing it to the actual fire count for validation.
By creating a system that blends machine learning with nearly real-time satellite measurements to estimate fire numbers themselves per 0.25-degree zone, we hope to do away with all of these restrictions while simultaneously producing a similar and quantifiable metric. Additionally, the model goes through continual ground truth validation to evaluate its performance. Our methodology is unique because it provides the count of probable fires for the next seven days on a weekly rolling basis- a quantity anyone can easily verify and validate. Our solution is much more dynamic than the cumbersome mathematical equations approach as it uses cutting-edge machine learning models.
Our Fire Prediction Model Output
Blue Sky Analytics’ Fire Predictions dataset provides the count of fires as the output. The plots below show India's Actual and Predicted Fire Events from 9th to 15th November 2020.
Revolutionising Fire Management Efforts
Our Fire Prediction dataset can empower various stakeholders, such as :
Municipal corporations: can use fire prediction data to identify high-risk areas for fires and take preventative measures to reduce the likelihood of fires occurring.
Utility companies: can use fire prediction data to inform the placement of equipment and infrastructure, such as power lines, to minimise the risk of fires.
Municipal bond traders: fire prediction data can be used to identify municipalities at high risk of fires and may have higher costs associated with fire prevention and response. This information can inform the traders' decisions about buying or selling bonds issued by those municipalities.
Insurance firms: fire prediction data allows insurance companies to identify specific areas, types of properties or businesses at high risk of fire and adjust pricing appropriately, in addition to helping in the underwriting process.
Learn more about our fire prediction dataset here.