A PFM/SHM-Aware Spatiotemporal Contextual Fire Detection and Adaptive Thresholding Framework for VIIRS 375 m Data

Thermal contextual algorithms for 375 m VIIRS active fire detection can produce substantial commission errors over persistent non-wildfire heat sources (e.g., refineries, gas flares, and volcanoes), and globally fixed thresholds may be suboptimal under heterogeneous thermal backgrounds. We present a lightweight spatiotemporal prior layer that augments by applying prior-guided, pixel-level parameter switching during the discrimination stage. The layer combines: (i) a persistent non-wildfire thermal anomaly mask (PFM) derived from multi-year VNP14IMG recurrence and seasonality statistics on a 0.004° grid, and (ii) a short-term heat-source mask (SHM) based on nighttime VIIRS I4/I5 brightness temperature stability to capture newly emerged or rapidly intensifying static sources. Pixels flagged by either prior are processed with a stricter parameter set, while other pixels follow the baseline setting. We evaluate the method using a stratified validation dataset (N = 3435) spanning industrial/urban clusters, volcanic regions, forest/grassland wildfires, and fragmented crop residue burning, with validation supported by independent high-resolution imagery (Sentinel-2/Landsat) and external POI datasets. The framework markedly reduces false positives in high-interference zones (industrial/urban false positive rate from 88.6% to 22.7%; volcanic from 100.0% to 57.3%) while preserving high performance for forest/grassland wildfires (F1 ≈ 0.999). For fragmented residue burning, omission error decreases from 11.2% to 1.3%, improving detection completeness without an apparent increase in commission errors. Overall, the results suggest that integrating long- and short-term spatiotemporal priors via threshold switching can improve the robustness and interpretability of contextual VIIRS fire detection under complex thermal backgrounds in the evaluated scenarios.