Geospatial System for Wildfire Monitoring and Prediction Using Aerospace Data

A unified wildfire early-warning system is assessed, linking Machine Learning (ML) detection with an observation-updated Cellular Automata (CA) spread model, relying only on data available at the time of issue (t 0 ). Inputs include satellite imagery, weather, and topography. The detector provides fire probability and Fire Radiative Power (FRP). The CA model is adjusted based on available detections and perimeters. Uncertainty is handled through an ensemble of model runs and isotonic calibration to adjust probabilities. The system is assessed using a forward-looking evaluation. Detection achieves an F1-score of 97.1%±0.3% and high Average Precision. The 24-hour spread forecast shows high accuracy with a mean IoU of 0.84±0.02 and a low Arrival time MAE of 21±3. The probabilistic forecasts demonstrate excellent reliability, with a Brier Score of 0.14±0.02 and an ECE of 0.028±0.006. The system operates with a median end to end latency of 300 minutes, supporting the required 6-hour issuance cadence. The system is compared against ML-only and physics-based models. Results show high detection accuracy and speed with low false alarms. Spread predictions closely match real perimeters and arrival times. Probability estimates are reliable. Early warnings typically give six or more hours of notice. The system runs efficiently, delivering results in about five hours. Using only t 0 inputs, observation-based updates, and calibrated probabilities yields accurate, easy-to-understand outputs for incident response and WUI planning. The alert threshold (τ) is adjustable, and frequent updates mitigate limits like cloud cover.