Tools to forecast fire spreading behavior

Hour-by-hour machine learning assisted forecast of forest fires shape and area

Okanagan Drone — OpenAir UAV Flight Training Partner

In the past year alone, the world has been witness to a worrying burst of forest fires: bushfires in Australia and the Amazon gained notoriety in late 2019, but other areas like Central Africa and Patagonia have large forest fires every year. Most recently, wild fires have been raving throughout almost every region of the world: Bolivia, Algeria, Canada, Brazil and North-eastern Argentina, most of Western Europe, Turkey, Russia and the list sadly goes on and on.

Forest fires are sadly here to stay, and the best and only thing we can do is figure out how to fight them.

Developed between November 2019 and February 2020, ForestFires.tools combines UAVs, embedded computing boards, visible spectrum imagery, machine learning classification of vegetation types and fire-frontier detection, in a mobile frontier mathematical model to give first responders the resources they need to fight large scale forest fires at an affordable cost.

An unusual solution

This set of fire-fighting tools was developed for the Unusual Solutions Competition, a global competition organized by WeRobotics, that invites individuals and teams from all over the planet to submit solutions on how to best tackle social, environmental and scientific challenges with technology.

The goal of the project was to aid fire responders in the planning of their efforts. But also, considering that many of the countries with the largest forest areas are in the global south, and are collectively less wealthy than their northern counterparts, we focused on putting together a solution that has a cost of around 600 USD, making it affordable to as many groups as possible.

In future iterations, we plan to work with local UAV groups and maker-spaces in all countries, so that the product, training and support can be offered at an even lower cost.

How it works

There are 5 key elements that make this solution different to anything else out there:

1. Multispectral imagery updated often with the use of UAVs
2. Machine learning algorithms: one to classify the terrain according to vegetation, and one to detect the frontier of the fire
3. A small, purpose built, portable weather station
4. Mathematical modeling of the fire
5. The solution works offline, generating near real-time forecast

All of these components combined enable us to build an accurate, local and almost real-time forecast of wildfire spreading.

By combining multispectral aerial images collected by drones with meteorological variables obtained by the portable station, like temperature, humidity, wind speed and direction, ForestFires.tools can create a more accurate, local, near real-time fires forecast, that allows firefighters to better understand and respond to forest fires.

The data sources gathered by the UAVs are then processed by an inexpensive, portable computer that doesn’t require an Internet connection to prepare a fast fire model capable of forecasting the spread of the fire. This is a key feature of the product: it works locally and offline to give fire responders maximum value with minimum constraints.

Applying machine learning algorithms to the imagery, the model classifies vegetation according to type (grass, trees, water) and to current state of dryness/humidity. With this classification plus experimental data from other researchers, we were able to gather the necessary data to integrate evaluations such as conductivity, thermal diffusion or vegetation density into our model. This imagery also allows us to detect the coordinates of the fire sources that will later become our initial conditions.

Combining all of that data in a finite element model with a moving boundary of the heat equation, we can forecast the fire spreading in the short term, which provides responders with more accurate and real-time information for decision-making. This gives firefighters autonomy on the field, without losing access to accurate, timely information.

Equation used for fire modeling, where each term (p, k, T) corresponds to a matrix obtained from the image processing, modified by climate conditions (w)

Approach & components

The toolset allows first responders to better understand where and how fast fire will spread, and use that information to make better decisions on resource allocation.

Planning for firefighting is based on a triangular model: topography - fuel -weather

After validating the need and approach with potential users, we focused on reducing uncertainty around the technical feasibility of the solution. For that we built several small prototypes of the key components:

1. Fire boundary detection: We used TensorFlow Object Detection API to train a machine learning model that detects the coordinates of fire spots. We obtained training and validation images from multiple videos of forest fires captured by drones. With only 30 images in the training set (which we tagged manually), we obtained an 83% average precision on fire coordinates detection. As the basis for the model, we used a pre-trained COCO-faster-rcnn and ran the training on a Nvidia Tesla K80.
2. Classification of ground coverage: we applied NDVI, an algorithm that enables the detection of vegetated areas and other types of surfaces, such as rock, water and soil. More importantly, NDVI allows us to classify how green or dry the vegetation is, which improves the accuracy of the fire spreading forecasts. Running an NDVI analysis requires aerial imagery captured with an infrared camera.
3. Spread forecasting: we started from the basic heat equation in 2D and iterated to incorporate additional parameters into it:
   a. Initial coordinates of fire spots.
   b. A map of underlying types of ground cover that determine the combustibility of the material.
   c. Weather conditions
   That model enabled us to forecast the action of the wind in the fire spreading and how the materials burn, predicting the final shape and area of a fire over specific periods of time.
   We also optimized the model to be run on a GPU (we are planning on using an Nvidia Jetson), so that we can provide an offline, affordable and portable platform to run the forecast.
4. Components for the weather station: we researched and evaluated different components for the weather station and optimization box, and came up with a minimum viable budget for it. The alternatives are quite promising, with viable options starting at a cost of 240 USD.

We leaned towards using Raspberry PIs in combination with the Nvidia Jetson to run offline imagery processing. The imagery processed using a lightweight optimized AI algorithm for shadow analysis can be used to build a Digital Elevation Model, a key component in the analysis of the spread of forest fires. Assessment of fuel moisture could also be done with similar methods.

In addition to that, information from a portable weather station provides temperature readings, humidity and wind speed and direction. That component is essential, partly because lack of connectivity means no access to remote weather stations, but also because fires create their own nearby weather conditions. A fire propagation model would then use that input data to forecast potential scenarios. The Jetson board is a perfect fit for all these tasks and comes at an extremely reasonable cost of 99 USD.

For image processing purposes, we based our code on an integration or extension of OpenDroneMap. For the Weather Station we think the best approach is using and interconnecting with the Raspberry PI Weather Station, These guarantee we build on existing open source and open hardware efforts, significantly reducing development costs and increasing long-term sustainability.

The resulting toolset

These wildfire behavior forecasting tools deliver an hour-by-hour machine learning assisted forecast of forest fires shape and area. The outcome is precise and timely knowledge in the hands of firefighters, which significantly improves their ability to allocate resources and make decisions, as efficiently as possible.

The potential impact of this project is a more focused and efficient use of resources during forest fires response efforts. If responders have a better understanding of the likelihood of the fire spreading in certain directions and at certain speeds, efforts can be better focused, and evacuations –if necessary– could be planned with more time.

Next steps

ForestFires.tools was developed as a concept for the Unusual Solutions Competition, and has not yet been deployed in a real setting. However, it is as close to completion as possible without financing.

Our roadmap to a complete and successful implementation is as follows:

1. NDVI expansion
   The NDVI analysis provides a good basis to understand the type of ground coverage and is widely used to assess crop health in agriculture and forest coverage from satellite remote sensing. In order to provide more accurate fire behavior forecasts, we need to expand on that and use machine learning to classify types of vegetation. The input to such a machine learning classifier is the multispectral imagery (both infrared and visible imagery).
2. Real-world testing
   Testing the spreading model, first without weather input. Using real imagery captured from past fire events, both aerial or satellite, and infrared spectrum, we can run our model and compare its predictions with the actual spreading. Those cases can be used to iterate and improve the model until the results are accurate enough.
3. Portable weather station development
   The portable, local weather station needs further development and testing, and to incorporate the output to the model. While this whole solution can be used with an off-the-shelf weather station, that single component is what increases the total cost the most. To improve on affordability, we have researched a few alternatives of wind sensors that can be integrated into a small, portable and affordable weather station that can be connected to the proposed solution.
4. NDVI testing on well-known terrain
   Run further tests on NDVI analysis using multispectral imagery captured with UAVs and the subsequent machine learning classification on well known terrain. We want to evaluate and compare lower-end infrared cameras against more expensive ones to determine the most affordable feasible option.
5. Contrasting accuracy in real setting
   Join a fire event response and test the whole solution, without acting on the forecasts, but contrasting them with the actual progress of the fire. We have already got clearance to do this in Patagonia.

We are very confident about the potential of this project and the value these tools can deliver to fire responders all over the world. As forest fires keep decimating key resources in our ecosystems, there is much we can do, and this feels like a good place to start.