FASDD: An Open-access 100,000-level Flame and Smoke Detection Dataset for Deep Learning in Fire Detection

• DOI: 10.5194/essd-2022-394
• Publication Date: 2022-11-21T08:46:15Z
• Authors: Ming Wang; Liangcun Jiang; Peng Yue; Dayu Yu; Tianyu Tuo
• Keywords: Fire Detection; Benchmark (surveying)

Abstract. Deep learning methods driven by in situ video and remote sensing images have been used fire detection. The performance generalization of detection models, however, are restricted the limited number modality training datasets. A large-scale benchmark dataset covering complex varied scenarios is urgently needed. This work constructs a 100,000-level Flame Smoke Detection Dataset (FASDD) based on multi-source heterogeneous flame smoke images. To best our knowledge, FASDD currently most versatile comprehensive for It provides challenging to drive continuous evolution models. Additionally, we formulate unified workflow preprocessing, annotation quality control samples. Meanwhile, out-of-the-box annotations published four different formats deep models trained demonstrate potential value challenges localization. Extensive evaluations classical show that can achieve satisfactory results, especially YOLOv5x achieves nearly 80 % mAP@0.5 accuracy spanning two domains computer vision sensing. And application wildfire location demonstrates be recognizing monitoring forest fires. deployed simultaneously watchtowers, drones optical satellites build satellite-ground cooperative observation network, which provide an important reference suppression, victim escape, firefighter rescue government decision-making. available from Science Data Bank website at https://doi.org/10.57760/sciencedb.j00104.00103 (Wang et al., 2022).