Wildfires in the U.S. alone cause billions in damages annually, with suppression costs exceeding $4 billion per year. FireCastRL is an AI framework that predicts wildfire ignition using spatiotemporal data and deploys RL agents for intelligent helitack suppression in simulated environments. Developed by Shaurya Mathur, Shreyas Bellary Manjunath, Nitin Kulkarni, and Alina Vereshchaka at the University at Buffalo, this project aims to enable proactive response, reducing risks and resource waste. This work was accepted for publication as a paper at the IEEE International Conference on Machine Learning and Applications (ICMLA 2025).

Key Features

• Public Resources: Access our custom dataset (9.5 million samples), custom gym-compatible wildfire environment, live FireCastRL web-app and GitHub code.

• Ignition Forecasting: A hybrid CNN-BiLSTM model achieving up to 73.1% accuracy on unseen data, predicting risks from 75-day meteorological sequences.

• RL-Based Mitigation: Physics-informed simulations where agents learn optimal helitack deployment, minimizing fire spread and resource use.

• Actionable Threat Reports: A fire threat assessment report containing predicted ignition coordinates, burn trajectory simulations, and strategic response recommendations for emergency responders.

FireCastRL is more than just a prediction model, it's a complete, five-stage pipeline designed to transform wildfire management from a reactive to a proactive process. The framework integrates data acquisition, forecasting, AI-driven simulation, and reporting into a single, cohesive system.

Existing public datasets for wildfire analysis primarily focus on detecting active wildfires using satellite or aerial imagery. While essential for real-time monitoring, these resources are fundamentally reactive, meaning they help us see a fire that has already started. This leaves a critical gap in the ability to predict where the next fire might ignite.

Our dataset was built to fill this gap. Instead of imagery, the FireCastRL dataset is built on rich, historical meteorological and environmental data. We created it by combining over a decade of wildfire ignition reports from IRWIN with daily weather sequences from GRIDMET. This approach allows our AI models to learn the subtle, pre-ignition conditions that signal rising fire danger, such as long-term drought, sudden temperature spikes, or shifts in wind patterns. 

Key Differentiators:

• Focus on Prediction, Not Detection: While other datasets use satellite images to find active fires, ours uses weather and environmental data to forecast potential ignitions before they happen.

• Spatiotemporal Depth: With records structured as 75-day time series, our dataset provides the historical context needed to understand how environmental conditions evolve over time to create fire risk.

• Balanced for Machine Learning: We developed a novel three tier negative sampling framework to generate non-wildfire samples, creating a balanced and robust dataset ideal for training accurate classification models.

By shifting the focus from imagery of active fires to the underlying environmental drivers, the FireCastRL dataset empowers researchers to build the next generation of forecasting tools, helping us get ahead of wildfires, not just react to them. The final public dataset contains 9.5 million labeled records structured as 75-day time series, capturing both wildfire and non-wildfire events.

• Incident Collection: We started with 348,604 wildfire reports from IRWIN (2014–2025) and applied a spatiotemporal filter to isolate 50,720 unique ignition events, removing duplicates and flare-ups.

• Negative Sampling: Since the data only provides positive labels (ignitions), we developed a three-tier negative sampling framework (far, near, and yearly negatives) to generate 76,080 non-wildfire samples for a balanced dataset.

• Feature Extraction: For each of the 126,800 total samples, we extracted a 75-day temporal window (60 days prior, 15 days after) of daily data from GRIDMET. This captures both the lead-up conditions and early post-ignition meteorology.

Each sample includes 15 environmental features that allow the model to learn both long-term stress signals and short-term ignition triggers.

For instance, a prolonged drought (captured through vpd, fuel moisture, and evapotranspiration) increases baseline fire risk, while short-term factors like a sudden wind spike or a hot, dry afternoon (captured by wind speed, temperature, and solar radiation) can act as immediate ignition triggers. By incorporating this diverse set of environmental variables, our model can learn nuanced temporal patterns that precede wildfires, enabling early and accurate predictions that improve preparedness and mitigation efforts.

Access The Dataset

The complete dataset is publicly available on Kaggle to support further research in wildfire forecasting and disaster response. 

Kaggle URL: https://www.kaggle.com/datasets/firecastrl/us-wildfire-dataset

Model Overview

We designed a hybrid neural network that combines a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (Bi-LSTM) network. This architecture is built to capture both spatial correlations and temporal dynamics from the 75-day data sequences.

Architecture Breakdown

• Input Layer: Takes a 75-day tensor of 15 meteorological features.

• Convolutional Layers: Two Conv1D layers (32 and 64 filters) distill spatial features from the time-series data. This helps the model identify significant patterns within the 15 environmental variables at each time step.

• Bidirectional LSTM Layer: A Bi-LSTM with 128 units processes the sequence both forwards and backwards. This is crucial for capturing temporal dependencies, such as how a prolonged drop in humidity affects fire risk days later.

• Output Layer: A final dense layer with a sigmoid activation function outputs a single value, the probability of a wildfire igniting.

This hybrid approach allows the model to learn not just what conditions are dangerous, but when and in what sequence they become critical.

Training

The CNN-BiLSTM model was trained on our 9.5 million sample dataset. To address the inherent class imbalance between ignition and non-ignition events, we employed techniques like focal loss and class-balanced sampling.

Our model demonstrates state-of-the-art performance in wildfire ignition prediction. On a strict held-out test set (data from Jan–Apr 2025), our CNN-BiLSTM model achieved 73.1% accuracy, outperforming several baseline machine learning and deep learning models.

The model's robustness was validated against real-world events. Notably, it successfully predicted the Palisades wildfire (January 2025) with a 98.6% confidence score.

Although weather-driven patterns explain most lightning-caused fires, human activity now accounts for ~84% of U.S. wildfire ignitions and greatly expands the season and locations where fires can start. Because these events are decoupled from meteorological precursors, our model cannot reliably foresee them. This limitation impacted the model’s current performance and tends to miss or underestimate the likelihood of human-triggered ignitions.

As a result, some false negatives and lower recall are attributable to the absence of human activity signals in the model’s inputs. Addressing this gap is a primary focus of future work: by integrating spatial and temporal data reflecting population density, land use, utility corridors, and holiday activity patterns, as well as exploring advanced AI architectures that can combine these diverse inputs, we aim to make FireCastRL more robust and versatile for real-world fire management in a human-influenced environment.

The simulation environment is based on a Cellular Automata engine, embedded in a custom Gymnasium wrapper. This setup allows the fire to evolve step-by-step while the agent interacts with the physics-informed 3D environment. Simulation state transitions evolve over discrete time steps, incorporating terrain elevation, wind direction, and land cover characteristics.

Our simulation backend is adapted from the Concord Consortium’s wildfire model. The terrain is rendered as a 240×160 grid, layered with digital elevation maps and real-time land cover data.

We use a fire engine module based on Rothermel’s fire spread equations, adapted for standard grid-based wildfire simulators such as FARSITE.

Each land cover represents a vegetation type that includes surface area-to-volume ratio, dry fuel availability, vertical extent of combustibles, fuel density, and minimum moisture for combustion. These values are derived from Anderson’s 13 Fuel Models and Scott & Burgan’s 40 Fuel Models, and adjusted for our simulation.

To generate a realistic simulation environment, we construct a 3D terrain for each high-risk forecast . Using the Google Earth Engine, we source MODIS land cover, SRTM elevation data, and GRIDMET wind data.

The reward function is based on short-term suppression with long-term containment by considering both state-based and proximity-aware feedback.

• Positive rewards: extinguish burning cells; prevent fire expansion; maintain proximity to fire fronts.

• Negative rewards: fire growth or large burnt area; inaction or delayed intervention; hovering over burnt or inert terrain.

This encourages the intelligent suppression strategies such as dynamic circling of fire fronts, firebreak creation, and early response to high-risk areas.

We extract a condensed feature vector from the structured observations using a custom multi-stage feature extractor. First, the stacked environment grids are processed by three parallel CNN branches with different receptive fields to capture fire patterns at multiple spatial scales. The resulting feature maps are unified and then refined by a spatial attention module to focus the network on the most critical fire zones. Finally, the flattened spatial features are fed into a two-layer LSTM, which integrates temporal information by learning from the historical dynamics of the fire and the agent's past actions.

Additionally, we process helitack coordinates passed through a multi-layer perceptron (MLP) encoder and a binary flag indicating if our RL agent is over a burning grid cell.

Finally, we concatenate all extracted features into a unified vector. This fused representation is then passed through a 2-layer fully connected residual block to produce the latent feature vector that serves as the input to the RL policy and value networks.

We use the Proximal Policy Optimization (PPO) algorithm to train our agent. PPO's stable learning process is well-suited for our dynamic wildfire suppression environment because it effectively handles complex state transitions influenced by agent actions and sparse rewards.

After 3X10^5 training steps, our PPO-based helitack agent reliably learned to circle the flame front, effectively slowing fire progression on held-out simulation maps. We compared this agent to a rule-based baseline that drops suppressant without observing the fire's state. 

The results show that combining an accurate forecaster with an RL agent enables effective suppression strategies, reduces projected damage, and provides a valuable benchmark dataset for the community.