-
Loss functions of discriminator and generator from CTGAN, generating synthetic data
-
Our website
-
Confusion matrix of CatBoost with synthetic data
Inspiration
Wildfires cause devastating damage when uncontrolled. The ability to predict wildfires in advance allows for effective prevention, better resource allocation, and faster response times, which can save lives and minimize destruction. The SAP Challenge introduced us to this pressing problem, and we were motivated to create a solution that helps people by providing tools for wildfire prediction, response, and damage mitigation. We envisioned a simple and accessible hub that could show historical fire data, track deployments, calculate costs, and even predict future fires to aid in relief efforts.
What it does
Hotzpot is an easy-to-use platform that visualizes wildfire data for Quebec, providing insights into fire locations, damage, costs, and more. Users can also upload their own data for visualization. Beyond displaying past fires, Hotzpot uses advanced AI algorithms powered by synthetic data generated through GenAI to predict future fires. The platform considers various factors, including the time of year, temperature, humidity, wind speed, vegetation, precipitation, and human activity.
Key outputs include:
Number of fires addressed: X Number of fires delayed: X Total operational costs: X Estimated damage costs from delayed responses: X Fire severity report: {'low': X, 'medium': X, 'high': X}
How we built it
Hotzpot is built using Django for the backend and React for the frontend, offering a responsive and interactive user interface.
Data Challenges: Our dataset was very small and unbalanced:
34931 instances of "no fire" 65 instances of "low fire" 38 instances of "medium fire" 30 instances of "high fire" With such an imbalanced dataset, training an accurate prediction model would be biased and ineffective. To tackle this challenge, we leveraged Generative Adversarial Networks (GANs), specifically CTGAN (Conditional Generative Adversarial Network), to generate synthetic data that simulates additional fire instances across all categories. This helped balance the dataset and improve model performance.
For the predictive model, we used CatBoost, a highly effective decision tree boosting algorithm known for its accuracy, especially on small and imbalanced datasets like ours. CatBoost has been shown to perform exceptionally well in wildfire prediction tasks.
Challenges we ran into
Unbalanced and small dataset: The biggest hurdle was working with a highly imbalanced dataset that contained very few examples of actual fires. Deployment issues with Azure: We faced difficulties integrating our application with Azure through GitHub, which delayed deployment. UI design: Striking a balance between a clean and intuitive UI while presenting a lot of data was tricky. Settling on a color palette: Finding the right color scheme that was both visually appealing and accessible took time.
Accomplishments that we're proud of
Successfully generating synthetic data using CTGAN, overcoming the small dataset challenge. Boosting the accuracy of our machine learning model and making it more reliable. Designing a fully functional platform that provides valuable predictions and data visualizations.
What we learned
How to work with frontend technologies like React and Django to create a seamless user experience. Strategies for handling unbalanced datasets in machine learning and how to use generative models to improve data quality. The importance of a user-friendly design when presenting complex data.
What's next for Hotzpot
Data Expansion: We'll scale up the dataset by collaborating with government and environmental organizations to collect more wildfire data, improving the model’s accuracy. Improved Predictions: Enhance the AI model to predict fire spread, intensity, and environmental impact using additional data points such as topography and real-time satellite imagery. Global Expansion: Extend Hotzpot’s coverage beyond Quebec, making it a global platform for wildfire prevention and response. Mobile Application: We plan to develop a mobile app for real-time access to wildfire data, notifications, and user uploads. Emergency Response Integration: Partner with emergency services to integrate real-time fire alerts into their systems for faster responses. Natural Disaster Expansion: Extend the platform’s capabilities to include predictions for other natural disasters like floods, hurricanes, and tornadoes. With Hotzpot, we aim to save lives, reduce damage, and make wildfire management more efficient, all while advancing the field of AI and environmental disaster response.
Log in or sign up for Devpost to join the conversation.