Accelerating Wildfire Spread Prediction for Guidance: A ConvLSTM Deep Learning Approach

Wildfire simulations play a critical role in emergency response planning, resource allocation, and the deployment of autonomous systems such as Unmanned Aerial Systems (UAS) for monitoring and suppression. However, physically accurate models for wildfire propagation are often too computationally expensive to be deployed in real-time applications, particularly those involving iterative decision-making like Reinforcement Learning (RL)-based guidance systems. In this work, we propose a fast deep learning-based model that approximates the dynamics of a wildfire propagation simulator. We use a Recurrent Neural Network (RNN) architecture based on Convolutional Long Short-Term Memory (ConvLSTM) units, trained on data generated from a deterministic Cellular Automata (CA) fire model. Our model learns to predict the fire state at the next time step, conditioned on the current state to simulate multiple time steps forward, producing the fire spread patterns with significantly reduced computational cost. The network is designed to capture long-range spatial–temporal dependencies that commonly drive nonlinear wildfire behavior. This enables the model to generalize across diverse ignition patterns and environmental conditions present in the training set. Results indicate good accuracy compared to the ground truth and up to two orders of magnitude faster propagation times, showing promise for high-frequency deployment in guidance loops.