Improving Wildfire Boundary Estimation through Temporal Multiscale Fusion of GOES-R and VIIRS Satellite Data

Abstract

Wildfire monitoring relies on remote sensing, yet no single sensor offers both the temporal and spatial resolution needed for fast fire detection and tracking fire spread in near-real-time. Geostationary sensors like Geostationary Operational Environmental Satellite-R (GOES-R) have low spatial resolution (~2km) but capture a scene every few minutes, while polar-orbiting sensors like Visible Infrared Imaging Radiometer Suite (VIIRS) have high spatial (375m) but revisit a site once or twice daily. Existing fusion approaches use only a single GOES image paired with VIIRS, overlooking the temporal sequence of fire growth recorded across GOES's frequent observations. We introduced a dual-encoder network that preserves this temporal context by processing 24-frame GOES-R sequences alongside single-pass VIIRS imagery. Each GOES encoder stage condenses its temporal stack into a multiscale summary tensor, then fused with VIIRS features through learned gating blocks at four resolution levels. A U-Net-style decoder then reconstructs the fullresolution fire-probability mask. We trained and evaluated the model with 5-fold cross-validation across 234 scenes from the 2021 Dixie Fire, reaching F1=0.751 (precision=0.708, recall=0.800, IoU=0.602, AP=0.816). On the hidden test set, a 5- model ensemble with test-time augmentation achieved F1=0.634 (IoU=0.465), a significant improvement over the earlier single-pathway architecture (F1=0.4375, IoU=0.280). Predictive uncertainty, estimated via Monte Carlo dropout, was low (mean entropy=0.0071, mutual information=0.0007), indicating confident and well-calibrated mask predictions. The results demonstrate that fusing GOES data with VIIRS imagery can produce sharper fire boundary estimates, though testing additional wildfires would be needed to confirm the approach generalizes.