P17 - Fire risk assessment using advanced sensing at the edge of drones
- Event
- iCCC2026 - iCampus Cottbus Conference
2026-05-05 - 2026-05-07
Cottbus - Band
- Poster
- Chapter
- Condition Monitoring / Predictive Maintenance
- Author(s)
- S. Devi - BTU Cottbus-Senftenberg, Cottbus, I. Jablonski - Fraunhofer IPMS, Dresden
- Pages
- 184 - 187
- DOI
- 10.5162/iCCC2026/P17
- ISBN
- 978-3-910600-10-2
- Price
- free
Abstract
Recent advancements in drone technologies have increased their applicability in image-based fire detection. This paper presents image-based fire classification using convolution-based autoencoders (CAEs) and extended fusion models such as early fusion, feature-based fusion, and decision-based fusion. This study provides an overview of using fusion algorithms to classify fire using the FLAME₃ dataset and a comparative study of the accuracy of the thermal dataset and the RGB dataset in the classification of fire. The results indicate that models trained on the thermal dataset could discriminate between Fire and No Fire instances better than models trained on the RGB dataset. Moreover, the CAE-based early fusion model achieved an accuracy around 95%, which is more as compared to feature level (92%) and decision level (91%). It can be concluded that fusion approaches are credible for fire classification and have the potential to localize fire-prone areas. Future studies aim to achieve the classification and localization of fire risks in mined-out areas in open-cast mining using different sensors on drone platforms. As a guide for future work, the focus would first be to reduce the complexity and size of AI-based models using pruning, quantization, and knowledge distillation so that they can be used on drone platforms to run real-time inference. As a collaborative edge-cloud-based approach, edge AI (artificial intelligence) servers are proposed to be to run these lightweight models onboard, and computationally expensive tasks and processing can be shifted to cloud platforms.