LIT | Devpost

Inspiration

Predicting wildfire spread can help save both lives and millions of dollars in damage by giving communities and emergency responders more time to prepare and react. In recent years, massive wildfires across Oregon and the western United States have shown how quickly fires can overwhelm towns, destroy ecosystems, and strain emergency response systems. Traditional wildfire modeling often relies on computationally expensive physics simulations that can take a long time to process, which becomes a major issue during fast moving disasters. We wanted to explore whether machine learning and interactive visualization could provide faster, more accessible predictions. That idea inspired us to build LIT, a system focused on generating real time, data driven wildfire spread simulations that are easy to visualize and interact with.

What it does

LIT is an interactive wildfire spread simulation platform that allows users to visualize how a fire may expand over time across Oregon. Users can select an ignition point directly on a map and observe predicted burn severity regions as the simulation updates. The platform displays fire spread visually through a grid overlay while also tracking important information such as estimated burn area, environmental conditions, and simulation progress. The goal of the project is to provide a fast, intuitive way to understand wildfire behavior and demonstrate how predictive systems could support disaster awareness and emergency planning.

How we built it

We built LIT using a Flask backend and a React frontend. The frontend handles the interactive visualization by displaying an Oregon map and rendering fire severity overlays onto a canvas based grid system. Users can click on cells to ignite fires or advance the simulation over time while dashboard components dynamically update important statistics. The backend manages the wildfire prediction pipeline. It exposes API routes that return Oregon grid metadata and fire prediction results. To generate predictions, the system processes geographic data such as elevation, wind conditions, and ignition location information. We trained a PyTorch convolutional neural network using a UNet style architecture on Oregon related wildfire examples. The model takes localized terrain and environmental inputs and predicts the probability of burned regions surrounding the ignition point. To improve reliability, we also implemented fallback systems using heuristic fire spread generation when trained model data or environmental datasets are unavailable. This allowed the application to continue functioning even in incomplete runtime environments.

Challenges we ran into

One of the biggest challenges was working with inconsistent and incomplete wildfire datasets. Environmental data such as elevation maps, wind conditions, and historical fire masks often came in different formats, resolutions, and coordinate systems, which required extensive preprocessing before training the model. Another challenge was balancing realism with speed. High accuracy wildfire simulations can become computationally expensive, so we needed to design a lightweight model that could still generate believable predictions quickly enough for an interactive application. We also faced difficulties integrating machine learning predictions into a live frontend visualization. Mapping sparse prediction outputs onto an interactive grid while maintaining smooth updates required careful optimization between the backend API and frontend rendering logic. Finally, creating reliable fallback systems was important because missing datasets or failed model loading could otherwise break the demo entirely.

Accomplishments that we're proud of

We are proud that we successfully created a fully interactive wildfire simulation platform that combines machine learning, environmental data, and real time visualization into a single experience. One accomplishment was training a functional CNN model capable of generating burn probability predictions using terrain and wind related inputs. We are also proud of building a responsive frontend that allows users to directly interact with the simulation and instantly visualize changing wildfire conditions. Another accomplishment was designing the system to remain functional even when environmental datasets or trained models are unavailable. By implementing fallback prediction systems, we ensured the demo could consistently produce visible wildfire spread behavior, making the application more stable and accessible for demonstrations.

What we learned

Through this project, we learned how challenging real world environmental modeling can be, especially when working with incomplete datasets and unpredictable natural systems. We gained experience training convolutional neural networks with geographic and spatial data while also learning how terrain and environmental conditions influence wildfire behavior. On the software side, we improved our skills in full stack development by connecting a Flask API with a React based visualization interface and managing communication between machine learning predictions and live frontend rendering. We also learned the importance of designing resilient systems. Building fallback prediction methods taught us that practical software must continue functioning even when ideal conditions or perfect data are unavailable.

What's next for LIT

In the future, we want to improve LIT by incorporating real time weather feeds, satellite imagery, and more detailed environmental datasets to increase prediction accuracy. We also want to expand the simulation beyond Oregon and support larger geographic regions with higher resolution terrain modeling. Another future goal is adding predictive evacuation and resource allocation tools that could help emergency responders identify high risk areas and prioritize aid more effectively. Long term, we envision LIT evolving from a demonstration project into a more advanced decision support platform that could assist with wildfire preparedness, public education, and disaster response planning.