Hydra is an innovative app designed to tackle the challenges wildfires pose. These natural disasters threaten both ecosystems and human communities, and in 2024 alone, the United States experienced 64,897 wildfires, scorching approximately 8.9 million acres, resulting in $275 billion of economic losses, the destruction of 16,200 structures, and untold human tragedy. While controlled burns mitigate risks and promote ecological health effectively, reducing the likelihood of a catastrophic wildfire by 68%, identifying the optimal locations for these controlled burns remains a complex task. Inspired by research like that of Coffield et al. (2022), which highlights the potential of machine learning to enhance wildfire prediction, our app integrates several environmental variables such as type and density of vegetation, soil moisture, wind speed and direction, temperature and humidity, topography with historical data into an intuitive platform that provides data-driven recommendations about the optimal location for a controlled burn within a given region. Not only does Hydra identify the optimal location for a controlled fire, but it also predicts its FRP (Fire Radiative Power) as a measure of intensity through an ML XGB Boost prediction model, allowing firefighters and forestry officials to understand whether a location is suitable for a controlled burn or if such a fire would be too intense to effectively manage. Furthermore, Hydra is a one-of-a-kind tool, unparalleled in the world of wildfire management. Hydra ultimately achieved a ~90.3% accuracy in identifying the optimal location for a controlled fire and serves as an extremely beneficial product.
We faced multiple challenges before we reached a desirable result. When developing the optimal location algorithm, we initially did not check whether the fire's location was in an industrialized area. This led to the identification of optimal fire locations near major roads or buildings, which was impractical. To address this, we implemented a function to calculate the number of industrialized features within a 1-mile radius to ensure the location of the controlled fire is not in an industrialized area. When training the ML model, we initially made API calls serially to download all the training datasets. This was extremely slow and would have taken multiple months to fetch the data. By using parallel calls and running our data fetcher on multiple AWS/EC2 machines, we were able to download all the training datasets within a week. However, because of numerous parallel calls, our calls were throttled by NASA LARC. We added a retry mechanism to address throttled calls, and we implemented an exponential backoff mechanism sleeping increasing intervals of 2x duration.