Spatiotemporal Wildfire Prediction and Reinforcement Learning for Helitack Suppression

AI summary

Overview

Research area: Applied machine learning for disaster response — combining spatiotemporal deep learning for wildfire ignition forecasting with reinforcement learning for aerial firefighting (helitack) suppression.

Technical level: Advanced. The paper assumes familiarity with CNNs, LSTMs, attention modules, Proximal Policy Optimization, and fire-spread physics (Rothermel's equations).

Scope: The paper presents FireCastRL, an end-to-end pipeline that predicts where wildfires will ignite, simulates suppression with an RL-controlled helitack agent in a physics-informed 3D environment, and produces a threat assessment report, alongside a public dataset of 9.5 million labeled environmental samples.

What This Paper Is About

Wildfire management in the U.S. is largely reactive: fires are fought only after they are detected, leaving little time for containment. The authors build a proactive system that first forecasts where and when a fire is likely to ignite using 75-day windows of meteorological and environmental data, then hands high-risk predictions to a reinforcement learning agent that learns how to deploy an aerial helitack unit in a physics-based fire simulation. The goal is a single pipeline that supports both advance forecasting and tactical suppression planning for emergency responders.

Key Contributions

1. A publicly released large-scale spatiotemporal dataset — 9.5 million labeled samples of environmental variables for wildfire prediction, derived from GRIDMET and IRWIN, structured as 75-day multivariate time series windows. Released on Kaggle.

2. A deep spatiotemporal ignition forecasting model — a hybrid CNN-LSTM binary classifier trained on 75-day windows to predict wildfire ignition from historical environmental patterns, reporting 73.1% accuracy on a held-out test set.

3. A physics-informed 3D wildfire simulation environment — a cellular automata engine on a 240 × 160 grid using real digital elevation maps, MODIS land cover, and GRIDMET mesoscale wind fields, adapted from the Concord Consortium's wildfire model and Rothermel's spread equations, wrapped as a Gymnasium environment.

4. An RL helitack suppression policy trained with PPO, plus an end-to-end pipeline that generates a fire threat assessment report containing predicted ignition coordinates, burn trajectory, suppression sequence, and response recommendations.

Main Findings

• Forecasting accuracy: The CNN-LSTM model achieved 73.1% accuracy on a held-out test set covering January 2025 to April 2025, with precision 0.71, recall 0.70, and F1 0.70.
• Palisades wildfire prediction: The model predicted the Palisades wildfire (January 2025) with 98.6% confidence.
• Confusion matrix: 536 true positives, 1251 false negatives, 1455 false positives, and 458 true negatives.
• Comparison to baselines: The CNN-LSTM outperformed all listed alternatives — XGBoost (66.4%), Gradient Boosting (65.6%), Random Forest (64.6%), K-Nearest Neighbors (63.7%), Simple-MLP (62.6%), Decision Tree (62.0%), Two-Layer-LSTM (61.9%), LightTS-Inspired (60.6%), Logistic Regression (59.7%), and Naive Bayes (56.5%). Non-time-series baselines were flattened to 1125 dimensions (15 variables × 75 timesteps).
• Error source: Performance was primarily limited by unpredictable, human-caused ignitions that are independent of weather patterns.
• PPO suppression performance: After 3 × 10⁵ training steps, the PPO agent learned to circle the flame front and slow fire progression on held-out simulation maps. It burned 1529 cells over 410 timesteps using 18 helitacks and 14,400 gal of water, versus the rule-based baseline's 4931 cells, 883 timesteps, 47 helitacks, and 37,600 gal of water — roughly a threefold reduction in burned cells and water use.
• Dataset composition: From 348,604 raw IRWIN reports (January 2014 to April 2025), filtering yielded 50,720 positive ignition events. Combined with 76,080 synthesized negatives (5,000 far, 35,000 near, 36,000 yearly), the labeled dataset totals 126,800 samples.

Methodology in Plain English

Building the dataset. The authors started with the IRWIN wildfire incident database and cleaned it: they kept only ignition events at least 5 km apart on any given date, required at least 2 hours between retained ignitions, and verified every coordinate fell inside the Continental U.S. bounding box (24.4°N–49.4°N, −125.0°W–−66.9°W). Because the database only labels where fires did happen, they generated three kinds of "no fire" examples: far-away points (5,000 coordinates at least 100 km from any fire), nearby points shifted 90–150 days in time (35,000 samples within 100 km of a real ignition), and the same locations one year earlier when no fire was recorded (36,000 samples). Each coordinate-date pair was expanded into a 75-day window — 60 days before and 15 days after ignition — with daily GRIDMET data at 4 km resolution, producing 15 environmental features (precipitation, humidity, wind speed, temperature, solar radiation, burning index, energy release component, fuel moisture, evapotranspiration, vapor pressure deficit, and others).

Forecasting. A hybrid CNN-LSTM treats ignition as binary classification over the 75-day sequence. Convolutional layers encode spatial structure; LSTM layers encode temporal evolution. Focal loss and class-balanced sampling counter the label imbalance. If the predicted probability exceeds a threshold ω, the system treats the location as high-risk.

Simulation. For high-risk forecasts, the system builds a 3D terrain from real elevation, land cover, and wind data sourced via Google Earth Engine (MODIS land cover, SRTM elevation, GRIDMET wind). The fire spreads step-by-step through a cellular automata engine using Rothermel's rate-of-spread equation, adjusted for wind and elevation effects, with fuel parameters drawn from Anderson's 13 Fuel Models and Scott & Burgan's 40 Fuel Models.

Reinforcement learning. A single helitack agent acts in a discrete action space (Up, Down, Left, Right, Drop). It observes a 4-frame stack of 160 × 240 environment grids showing fire states, its own grid position, and whether it is over a burning cell. Rewards are positive for extinguishing cells, preventing spread, and staying near fire fronts; negative for fire growth, inaction, and hovering over burnt terrain. The policy network combines three parallel CNN branches with different receptive fields, a spatial attention module, and a two-layer LSTM, feeding a two-layer fully connected residual block. Training used PPO with n_steps = 128, batch size 64, 3 epochs, learning rate 3 × 10⁻⁴, clip range 0.1, γ = 0.95, GAE-λ = 0.9, entropy coefficient 0.2, value-function coefficient 0.4, max-grad-norm 1.0, and target-KL 0.03.

Reporting. Each validation rollout logs suppressant drop coordinates and timestamps, burnt area trajectory, predicted ignition coordinates/time/confidence, deployment counts, time to containment, and suggested evacuation advisories — assembled into a threat assessment report, with a web application as the front end.

Why This Matters

Impact on research. The paper addresses a gap the authors identify explicitly: existing systems such as Google's FireSat, Canada's WildfireSat, and National FireGuard excel at detection and situational awareness, but do not handle downstream strategic response planning or autonomous suppression. FireCastRL argues that forecasting and mitigation belong in one closed loop, and its public release of 9.5 million labeled samples gives the community a benchmark for wildfire forecasting and AI-driven disaster response.

Real-world applications:

• Emergency responders receiving advance ignition-risk forecasts to pre-position crews and equipment before a fire starts.
• Tactical helitack planning — the simulation suggests where to drop suppressant and how much water a containment effort may require.
• Resource allocation across high-risk zones, prioritized through the threat assessment report's suppression prioritization zones and contingency thresholds.
• Evacuation advisory planning, using predicted burn trajectory and time-to-containment estimates.
• Policy planning for agencies balancing suppression budgets, given the reported $14.7 billion in direct property losses in 2023, over $3 billion in suppression spending, and an estimated total annual economic burden of up to $893 billion.

Industry relevance. The work sits at the intersection of geospatial AI, wildfire insurance and risk modeling, emergency management software, and aerial firefighting operations — all sectors that currently rely on reactive rather than predictive workflows.

Future Directions

1. Stronger forecasting architectures — the authors name Transformers as a direction for improving the ignition model beyond the CNN-LSTM.
2. Modeling human-caused ignitions — the reported error analysis points to unpredictable human activity as the main driver of forecasting failures, which weather-based features cannot capture.
3. Multi-agent suppression — extending the simulation from a single helitack unit to coordinated multi-agent setups.
4. Broader suppression tactics — incorporating ground crews and firebreak construction alongside aerial suppressant drops.

An additional open question the paper raises implicitly: the simulation omits several meteorological predictors used in forecasting (solar radiation and long-term humidity), and the geographic scope is restricted to CONUS, leaving generalization to other regions untested.

Target Audience

Machine learning researchers working on spatiotemporal forecasting, time-series classification, or RL for sequential decision-making; wildfire scientists and geospatial analysts interested in physics-informed simulation; emergency management and disaster-response practitioners evaluating predictive tooling; and policy makers concerned with suppression budgeting and resource allocation. Readers will benefit most if they have prior exposure to deep learning architectures and reinforcement learning, since the paper assumes that background rather than explaining it.

Authors’ abstract

Wildfires are growing in frequency and intensity, devastating ecosystems and communities while causing billions of dollars in suppression costs and economic damage annually in the U.S. Traditional wildfire management is mostly reactive, addressing fires only after they are detected. We introduce \textit{FireCastRL}, a proactive artificial intelligence (AI) framework that combines wildfire forecasting with intelligent suppression strategies. Our framework first uses a deep spatiotemporal model to predict wildfire ignition. For high-risk predictions, we deploy a pre-trained reinforcement learning (RL) agent to execute real-time suppression tactics with helitack units inside a physics-informed 3D simulation. The framework generates a threat assessment report to help emergency responders optimize resource allocation and planning. In addition, we are publicly releasing a large-scale, spatiotemporal dataset containing $\mathbf{9.5}$ million samples of environmental variables for wildfire prediction. Our work demonstrates how deep learning and RL can be combined to support both forecasting and tactical wildfire response. More details can be found at https://sites.google.com/view/firecastrl.