OT-WildfireState: A Multimodal Dataset for Short-Term Wildfire Spread Prediction

Existing wildfire spread benchmarks primarily target next-day prediction, providing limited support for short-term forecasting. We present OT-WildfireState, a multimodal benchmark for wildfire spread prediction over variable hourly horizons from 1 to 11 hours. Multi-satellite thermal-anomaly detections are aggregated into satellite-derived estimates of fire-affected areas, forming irregular sequences of wildfire states from which state-transition samples are constructed. The dataset contains approximately 15,000 transitions from 6,875 wildfire events across the United States and Canada, augmented with environmental and meteorological information. OT-WildfireState provides standardized splits, evaluation protocols, and baselines for reproducible short-term wildfire spread prediction.