Título: Deep Learning and Domain Adaptation for Wildfire Detection: Design Indicators for Scalable Camera -Monitoring Networks
Autores: Lucas Silveira, Ana Paula Batista & Everthon de Souza Oliveira
Resumo: Wildfires increasingly threaten ecosystems and public safety. This work presents a comprehensive wildfire-detection pipeline that combines a YOLOv11 backbone with Unsupervised Domain Adaptation (UDA) to deliver robust, real-time performance across diverse visual domains: we first curate and augment two public wildfire datasets (D-Fire and WildFire-Smoke-Dataset-YOLO) with mosaic, MixUp, geometric, photometric, and scaling transforms to simulate challenging conditions; then integrate adversarial feature alignment and pseudo-label self-training directly into YOLOv11 to align labeled source and unlabeled target feature distributions and mitigate performance drops under new lighting, vegetation, and camera settings. A systematic evaluation across five input resolutions (64–1024 px) shows that UDA raises a 640×640 model to 96.37% precision and a 1024×1024 model to 91.69% mAP@0.50:0.95—narrowing the accuracy gap with high-resolution baselines by over 40%. Finally, we derive a deployment-aware coverage model that links resolution, object size, and field-of-view to estimate required camera counts for any area, demonstrating that increasing resolution from 640 to 1024 px in a 10 ha scenario cuts hardware needs from 503 to 196 cameras (over 60% savings). Together, these contributions—domain-adapted detection, realistic augmentation, multi-resolution benchmarks, and quantitative deployment guidelines—offer a scalable, cost-effective framework for edge-capable wildfire monitoring.
Palavras-chave: Wildfire detection; YOLOv11; Domain adaptation; Real-time object detection; Camera deployment; Edge AI.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191745
Artigo em PDF: CBIC_2025_paper1191745.pdf
Arquivo BibTeX:
CBIC_2025_1191745.bib