Tehnički vjesnik, Vol. 33 No. 6, 2026.
Izvorni znanstveni članak
https://doi.org/10.17559/TV-20251028003098
Spatio-Temporal and Cross-Domain Learning Framework for Wildfire Detection in Power Transmission Corridors
Saipeng Zhang
; Jilin Electric Power Research Institute Co., Ltd., 130021 Changchun Jilin, China
*
Changlong Gao
; Jilin Electric Power Research Institute Co., Ltd., 130021 Changchun Jilin, China
Yufeng Tai
; Jilin Electric Power Research Institute Co., Ltd., 130021 Changchun Jilin, China
Junbo Liu
; State Grid Jilin Electric Power Company Limited Electric Power Research Institute, 130021 Changchun Jilin, China
Jingyao Luan
; Jilin Electric Power Research Institute Co., Ltd., 130021 Changchun Jilin, China
Chunshen Wang
; Jilin Electric Power Research Institute Co., Ltd., 130021 Changchun Jilin, China
Yufeng Liu
; State Grid Jilin Electric Power Company Limited Electric Power Research Institute, 130021 Changchun Jilin, China
* Dopisni autor.
Sažetak
Wildfires within power transmission corridors pose substantial risks to grid security due to their rapid spread, complex visual characteristics, and strong sensitivity to domain variations across terrain and environmental conditions. To address these challenges, this study proposes a wildfire detection framework that integrates spatio-temporal feature modelling with a cross-domain learning strategy. The framework employs optical-flow-guided Gaussian mask propagation for spatial enhancement and utilizes dual-level temporal consistency contrastive learning for dynamic behaviour modelling. A two-stage collaborative adaptation mechanism is adopted to align source-domain and target-domain representations through multi-model pseudo-labelling and semantic-guided feature fusion. Experiments conducted on a publicly available smoke dataset and a self-collected transmission corridor wildfire dataset showed that the proposed method achieved an accuracy of 94.5%, an F1 score of 94.4%, and an AUC-ROC of 0.978, outperforming multiple state-of-the-art domain adaptation baselines. The results demonstrate that the model effectively suppresses background interference, reduces false detections, and substantially enhances cross-scene generalization capability, offering a promising solution for reliable wildfire monitoring in complex power corridor environments.
Ključne riječi
cross-domain learning; semantic mask; spatio-temporal feature fusion; transmission corridor; wildfire detection
Hrčak ID:
353094
URI
Datum izdavanja:
7.11.2026.
Posjeta: 0 *