Phantom Fires: Safe False-Alert Suppression via Multimodal Fusion of a Visible-Patch CNN and Solar Reflection Geometry for Thermal Fault Monitoring

Abstract

Continuous outdoor thermal monitoring raises an analyst alert whenever an IR pixel exceeds a fixed apparent-temperature threshold. We report a 40-day deployment over 10 cameras (4,143.6 camera-hours) that produced 157 above-threshold events (≥ 150 °C): 122 (77.7 %) solar-glare false positives and 35 (22.3 %) confirmed true thermal events — a glare event rate of 0.0294 events per camera-hour (95 % CI [0.0157, 0.0449], bootstrap n = 1000). To suppress this false-alarm burden without compromising thermal-event recall, we construct a paired visible-patch dataset of 10,622 RGB patches and evaluate a ladder of five glare/noglare discriminators of increasing complexity, from a brightness threshold through a forward Sandia SGHAT geometric model, a logistic regression over 28 hand-crafted features, a small CNN (245,889 parameters), and a multimodal late fusion of the CNN and geometric scores. The fusion classifier reaches 98.6 % accuracy on a balanced 2,126-patch held-out eval set, a 3.3-percentage-point gain over the CNN alone and a 14.1-point gain over the geometric model; the learned weights (w_cnn = 7.34, w_geo = 4.50, bias = −1.93) confirm both streams carry independent signal, and adding the geometric stream cuts the CNN's false-positive count by 73 % while preserving recall. A separate system-level test on the 35 in-window true thermal events shows that both the CNN and the fusion classifier reach 99.58 % specificity (235 of 236 IR/VIS frame pairs preserved at the 150 °C production threshold). The current production deployment runs no glare suppression in place — every above-threshold IR event raises an analyst alert today — so the five-rung ladder and the multimodal fusion are research-grade results, not deployed artefacts. We propose a three-way routing policy (alert, fallback, suppress) that wraps the best rung and report the measured anchors that constrain its expected outcome on the deployment ledger.