FS-YOLOv3: A Reliability-Driven, Temporally Consistent, and Scene-Adaptive Dual-Source Forest Smoke Detector

Early smoke detection for forest fire prevention requires accurate and temporally stable decisions under dynamic clutter, tiny long-range targets, atmospheric degradation, and partial sensor unreliability. This paper presents FS-YOLOv3, a reliability-driven RGB–thermal smoke detector that extends a reproduced FS-YOLO baseline with two new modules: Cross-Temporal Consistency Alignment (CTCA) and Scene-Adaptive Expert Routing Fusion (SAERF). CTCA performs local short-horizon feature alignment and is evaluated with additional offset-field diagnostics to test whether the learned offsets correlate more strongly with annotated smoke expansion than with non-smoke motion. SAERF routes fused features to compact experts according to illumination, haze, texture ambiguity, and thermal reliability, with descriptor ablations and collinearity diagnostics used to examine routing stability. On the proposed clip-level RGB–thermal benchmark, FS-YOLOv3 improves over the reproduced FS-YOLO baseline from 93.7% to 96.3% [email protected] and from 89.5% to 94.8% temporal alarm consistency (TAC), with 165 model FPS on Jetson AGX Orin under the default one-frame-look-ahead buffered inference setting. Comparisons with lightweight YOLO detectors, RGB-only and infrared-only controls, simple fusion strategies, and stronger temporal baselines provide deployment context, while the main technical evidence is the controlled gain obtained by enabling CTCA and SAERF on the same baseline architecture. To support reproducibility, the paper specifies the baseline interface, sensor and annotation protocol, sequence-disjoint split policy, temporal metrics, threshold sensitivity, causal CTCA behavior, SAERF descriptor analysis, and model-side versus end-to-end latency boundaries. The reproducibility package is organized to provide code, configuration files, split identifiers, evaluation scripts, diagnostic-statistic scripts, and illustrative sample annotations; redistribution of the full curated benchmark is handled through institutional data-review approval or controlled access when direct video release is restricted.