Near-Real-Time Wildfire Segmentation and Brightness Temperature Estimation from Geostationary Satellite Imagery Using Deep Learning

Monitoring the progression of wildfires in near-real-time is essential for active-fire situational awareness and emergency response management. Current satellite-based wildfire monitoring systems face a trade-off between temporal and spatial resolution. Geostationary satellites such as the Geostationary Operational Environmental Satellite (GOES) offer frequent but coarse observations (~5–15 min, 2 km), while polar-orbiting satellites such as those carrying the Visible Infrared Imaging Radiometer Suite (VIIRS) provide fine spatial detail with limited temporal coverage (~12 h, 375 m). To bridge this gap, this study introduces a deep learning (DL) approach for sub-pixel fire segmentation and brightness temperature (BT) estimation from GOES imagery. The proposed approach consists of two steps, a segmentation step to distinguish active fire regions from background and a regression step to estimate the BT of active fire pixels. The model is developed using a dataset of 257 wildfires across the United States, consisting of multi-spectral GOES imagery paired with VIIRS-derived fire observations processed through parallax correction, reprojection, and resampling. The proposed approach processes data within approximately 5 min of observation, indicating near-real-time capability. Compared with a previously developed autoencoder-based model, the proposed method improves fire segmentation performance, increasing the average precision (AP) from 0.105 to 0.368 and reducing the fire-region root mean square error (RMSE) of BT estimation by 72.1% from 81.33 K to 22.69 K. Notably, the proposed approach reduces the omission error rate for medium-sized fires from 9.41% to 1.96% while maintaining zero omission errors for large fires. It also reduces the false positive rate on non-fire images from 66.86% to 8.48%. Ablation studies assess the impact of background values in the VIIRS reference data as well as the impact of loss functions and their weights. Overall, the proposed approach provides a practical solution for near-real-time sub-pixel fire segmentation and BT estimation using existing geostationary satellite systems.