EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Wildfire Fire-Smoke Detection in UAV Imagery via Scale-Guided Feature Routing and Uncertainty-Aware Localization

Linyu Wu

IEEE Access · 2026 · 14 · 96854-96876

DOI: 10.1109/access.2026.3706370 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Linyu Wu. Die Arbeit nennt 64 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Wildfire fire-smoke detection in UAV imagery is highly challenging because tiny early-stage fire sources and large diffuse smoke regions often coexist in the same scene, resulting in severe cross-target scale discrepancy. Such a characteristic makes it difficult for conventional detectors to simultaneously achieve robust multi-scale representation and stable localization. To address this practical problem, this paper presents WFSDet, a scale-aware extension to YOLO11 for wildfire fire-smoke detection in UAV imagery. WFSDet introduces two task-oriented designs to address cross-target scale discrepancy. First, the Scale-Guided Feature Routing (SGFR) module performs target-level-conditioned cross-level feature routing among multi-level pyramid features, enabling the detector to preserve fine-grained cues for small fire sources while retaining semantic context for large smoke regions. Second, the Scale-Discrepancy Penalized Regression (SDPR) objective explicitly models relative scale mismatch between predicted boxes and ground-truth boxes in log-scale space, and its uncertainty-aware extension further adaptively modulates scale-related localization according to sample-level uncertainty. Extensive experiments on the DFS, M4SFWD, and YUXI datasets show that the proposed method consistently improves detection accuracy, especially in scenes with severe target-scale variation. Specifically, WFSDet achieves $\mathrm {mAP}_{50}$ of 0.642, 0.943, and 0.775 on DFS, M4SFWD, and YUXI, respectively. TensorRT deployment on an NVIDIA Jetson Orin NX platform achieves 18.2 FPS, further demonstrating its practical feasibility for edge-side UAV wildfire monitoring. These results indicate that WFSDet provides an effective real-time-capable targeted solution for robust wildfire fire-smoke detection under severe scale discrepancy.

Zitieren

APA

Wu, L. (2026). Wildfire Fire-Smoke Detection in UAV Imagery via Scale-Guided Feature Routing and Uncertainty-Aware Localization. IEEE Access, 14, 96854-96876. https://doi.org/10.1109/access.2026.3706370

Vancouver

Wu L. Wildfire Fire-Smoke Detection in UAV Imagery via Scale-Guided Feature Routing and Uncertainty-Aware Localization. IEEE Access. 2026;14:96854-96876. doi:10.1109/access.2026.3706370

BibTeX
@article{Wu2026,
  author = {Wu, Linyu},
  title = {Wildfire Fire-Smoke Detection in UAV Imagery via Scale-Guided Feature Routing and Uncertainty-Aware Localization},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {96854--96876},
  doi = {10.1109/access.2026.3706370}
}
Kostenlos starten →