Subpixel Low-Intensity Fire Detection in Wide-Swath Remote Sensing: Dual-Scale Thermal Background Adaptation and Spatiotemporal Semantic Constraints

Subpixel low-intensity fires are difficult to detect in wide-swath imagery because mixed pixels weaken their thermal signals, whereas industrial facilities and photovoltaic panels can generate false alarms. We propose a dual-scale adaptive thermal background thresholding method based on spatiotemporal semantic constraints and conduct experiments using the detection of crop straw burning in China as a specific application scenario. The method combines scene-wide thermal state classification with local neighborhood statistics to adjust the above-background threshold and recover additional weak fire candidates. Farmland support, a persistent heat source prior derived from long-term VIIRS observations, and candidate-to-source proximity are then applied to newly added candidates to suppress non-target anomalies while preserving strong thermal detections. Experiments using 36 Wide Swath Imager scenes, near-contemporaneous Visible Infrared Imaging Radiometer Suite observations, field surveys in Anhui Province, and 260 Moderate Resolution Imaging Spectroradiometer scenes support the method’s use for weak fire candidate screening under heterogeneous thermal backgrounds. Targeted field surveys found fresh crop residue burn marks at 10 of 12 selected T0 candidate sites (83.3%). This proportion applies only to the surveyed candidate set and does not represent an estimate of overall detection accuracy. In the MODIS transfer experiment, detections near industrial facilities decreased from 45.17% for the official products to 7.92% with the transferred method. The proposed framework supports the regional screening of low-intensity fire candidates and targeted field verification, with persistent heat source information providing contextual constraints.