FireForesight

Wildfires pose a significant threat to human life and property globally, causing significant death and destruction each year. In California alone, the 2018 wildfire season led to almost 2 million acres of burned land along with over $26 billion in property damages. These wildfires affect communities worldwide, with around 29 million Americans being at risk.  As climate change compounds, wildfires are expected to worsen both in frequency and severity. Due to these factors, research aimed at developing technologies to fight wildfires has been at the forefront of environmental efforts. However, while current technologies to detect wildfires once they have started are very established, preventative methods which seek to stop wildfires from starting in the first place have been a lot more lacking. More specifically, few technologies exist that provide insights for where wildfires are most likely to happen. As such, firefighters cannot effectively focus their efforts within specific locations. Instead, they must divide their valuable time and resources across vast areas, leading to critical oversights in many high-risk locations. This not only impacts firefighting agencies, but also infrastructure dependent private companies. The very little amount of research that has been conducted regarding predicting wildfires has exclusively relied on the use of satellite flyovers for imaging. Due to the rapid nature of wildfire spread, satellites that can sometimes take multiple weeks to orbit over a certain location are often too late to allow any firefighters to perform preventative measures. Furthermore, satellite imagery has to be processed and analyzed offsite, further increasing latency in acquiring predictive insights. Many satellites currently in orbit are also decades old and thus have outdated sensors that cannot provide high resolution optical and thermal imagery. The lack of accurate wildfire spatial risk prediction tools contributes to their destructive potential, which we hope to address through our solution utilizing drone imagery and geospatial machine learning.  

Our innovation addresses the gap in preventative tools for wildfires. Through our AI-based solution, firefighters will be able to accurately anticipate which areas are most at risk for wildfires and thus focus their preventative efforts efficiently. The concept comprises of a machine learning software component and a drone platform hardware component.

For the software component, we assembled a massive geospatial dataset of optical and infrared images of the United States acquired by the United States Department of Agriculture National Agricultural Imagery Program (NAIP). These aerial images differ from previously used satellite imagery such as the Sentinel-2 collection in that they are conducted by relatively low flying aircraft with powerful cameras. This leads to a tenfold increase in resolution, greatly increasing the accuracy of our predictions. Using these aerial images, we can measure vegetation health and thus burnability through calculating the Normalized Difference Vegetation Index (NDVI) values for each image . This data is then used to train a neural network, which is then saved and exported to be deployed onto the drone. While in operation, the drone periodically images an area while saving the location of each image. The AI model runs completely onboard the Jetson Nano, allowing for real-time analytics. The drone transmits a map of the area it surveys with each area represented by a wildfire risk index.

Our custom drone platform pairs a NVIDIA Jetson Nano AI processor with an array of optical and thermal cameras. Our current iteration of the drone uses a hexacopter layout in order to increase range and payload capacity, though the drone itself can be customized in order to meet the needs of firefighting agencies. We mounted a short-wave infrared and 8MP visible spectrum camera and connected it to the Jetson Nano. Furthermore, we currently package the Jetson Nano and sensors as a detachable module, allowing firefighting agencies to utilize their own drone platform should they have more specialized needs. The drone is designed to be iterative; we plan on implementing more sensors including LiDAR scanners, anemometers, and soil moisture scanners while also updating our accompanying machine learning model to accept these inputs.

This geospatial wildfire risk prediction platform seeks to serve those living in areas with high wildfire risk. Because of the current lack of such prediction tools, firefighters in these areas susceptible to wildfires do not have location specialization, preventing them from focusing their efforts on high risk location, thus hindering their ability to fight fires. FireForesight would allow them to more effectively prevent the actual outbreak of the wildfires, preventing significant loss of life and destruction of property.

As our team is based in Orange County, California, we grew up experiencing firsthand the immense destructive potential of wildfires. During the 2020 wildfire season, the Silverado Fires reached mere yards from the community in which our team leader resided in. As a result, our team is especially prepared to develop and implement our drone platform solution, using the insights we have acquired from directly experiencing increasingly destructive wildfire season in California throughout the years. Furthermore, because of our location in an area historically prone to wildfires, there are many expansive county and state level firefighting agencies. This would allow us to closely work with a multitude of agencies, providing us with region-specific needs and insights.

We have developed promising prototypes in both the software and hardware side. For software, we have trained two preliminary models, each with different data features as input. The first model uses patches of Normalized Difference Vegetation Index (NDVI) aerial images as an input, achieving an accuracy of 100% evaluated through cross-validation with a test dataset of 350 randomly sampled datapoints. The second model combines these NDVI patches with topography and land surface temperature data in order to improve model generalization to varying global vegetation landscapes, achieving an accuracy of 99.71% through the same cross-validation testing as the first model.

Currently, our models rely on open-source, publicly available geospatial datasets. We hope that Solve can connect us with data analytics partners in order to help us acquire access to better, oftentimes exclusive data in order to further refine our geospatial machine learning models. Furthermore, we hope that Solve can also allow us to acquire advanced sensors for our physical drone platform, such as LiDAR scanners, drone-based anemometers, and soil moisture scanners in order to improve the accuracy of our predictions. We also are seeking internal management and outreach guidance from Solve.

Our innovation's reliance on drones aims to resolve the shortcomings of using outdated satellite imagery to predict wildfires. When it comes to wildfire prediction, punctuality is key. Wildfires can start burning in an instant and spread at blistering speeds. UAVs can be deployed within minutes and access difficult terrain through airborne travel while still covering large areas. Compared to satellites, which have fixed orbits, this represents a massive leap in the time it takes to analyze a given area. This advantage is further compounded by the fact that the entire machine learning package is stored locally within the Jetson Nano processor, allowing for real-time predictions. This cuts out the timely step of offloading satellite imagery to be analyzed offsite, greatly reducing the delay it takes to acquire predictions. Drones can provide much higher spatial imaging due to their low flight altitude, further increasing the accuracy of wildfire risk predictions compared to satellites. Satellites in orbit cannot be upgraded in a cost-efficient manner, resulting in sensors becoming outdated quickly. In contrast, our drone platform is, by nature, modular, allowing individual agencies to add on their own sensors for region-specific needs while also updating them constantly to match the state-of-the-art.

Through using our drone platform, firefighters will be able to derive near-instantaneous risk insights that are also much more accurate than traditional satellite imagery due to the massive leap in spatial resolution. As a result, they will be able to focus their preventative measures more precisely, stopping many fires from starting in high risk areas. This would greatly decrease the destructive potential of annual wildfire seasons. 

We hope that we can effectively deploy our drone platform initially with local agencies such as the Orange County Fire Authority. Once more thorough testing and refinement has been completed, our next goal would be to deploy our platform with state-level agencies, such as CalFire.

The core technology that powers our solution is geospatial machine learning, and more specifically, the recent research that has revolved around convolutional neural networks for image classification. Drones also form the backbone of our hardware components, with a variety of high resolution sensors being used in order to collect input data for the machine learning model.

I am currently the sole part-time engineer on our solution team, as I am also a student. 

I have been working on the solution for around 6 months now.

Our primary business model would be the business-to-government (B2G) model since our primary market consists of local and state level firefighting agencies. We hope to primarily provide the software service in the form of the machine learning model to them. Our secondary business model would be a subscription-based model for infrastructure dependent private companies who wish to utilize our platform. We would charge a one time fee for any hardware they require, and then offer the machine learning model as a subscription service.

Currently, we have raised around ~$200,000 through seed investment, giving us the ability to develop the initial models of the drone platform. However, for future iterations, we anticipate that we will receive significant government funding through subsidies promoting technological solutions to pressing environmental and social issues such as wildfire prevention. Our goals align with many government initiatives such as the NIH Climate Change and Health Initiative. We will also have a per-copy fee for private companies, affording us steady cash flow should government funding be inconsistent. Through utilizing cloud computing, we have reduced our computing costs for training our machine learning to effectively zero. Due to the lightweight nature of our algorithm, the cost of digital distribution would also be negligible. Our main cost would be the physical hardware, which we would have to supply to private corporations. Because firefighting agencies already have their own drone systems in place, our cost to deliver one service unit to them will be negligible. In the case of private companies, partnering with an established drone company will allow us to charge a premium as the middle-man and earn from both the software and hardware delivery.