Thesis
EARLY WILDFIRE DETECTION VIA UAVS: MANAGING DATA SCARCITY AND RESOURCE CONSTRAINTS
Washington State University
Master of Science (MS), Washington State University
05/2025
DOI:
https://doi.org/10.7273/000007384
Abstract
In recent times, climate change is an increasingly important environmental issue. One of the side effects of rising temperature is a significant increase in wildfires. The
resulting wildfires cause a significant amount of damage ranging from direct property damage to decreases in air quality. Due to the unexpected nature of the event, wildfire
management is a difficult task. The best method for managing wildfires is through early wildfire detection. Traditional methods for early wildfire detection rely on either
fire watchtowers or satellites. However, these methods suffer in either the temporal or spatial resolution.
Recently, the use of Unmanned Aerial Vehicles (UAVs) combined with on-board deep learning has received a surge of interest for early wildfire detection. Compared to traditional methods, UAVs have improved spatial and temporal resolution while being cheaper to implement. However, current research into the solution still suffers in several key aspects. The deep learning can be categorized into training, where a model is trained to detect wildfires, and inference, where the model is deployed on-board. For the training process, data collection and generalization of models have been identified as key issues. To address the issue of data collection, the use of One Class Classification is proposed. Various One Class Models were evaluated on the FLAME dataset and achieved an accuracy of 92%. For inference, the on-board nature of the application makes resource constraints a major design problem. In particular, energy usage, memory usage and inference speed are important metrics to consider. To improve the energy usage and inference speed of an on-board model, a method using sequential pixel variation is proposed. The method was evaluated on the FLAME dataset and achieved significant reductions in energy usage and inference speed while maintaining similar accuracy.
Metrics
1 File views/ downloads
17 Record Views
Details
- Title
- EARLY WILDFIRE DETECTION VIA UAVS
- Creators
- Wen Le Hong
- Contributors
- Xinghui Zhao (Chair)Josue Campos do Prado (Committee Member)Ji Yun Lee (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- School of Engineering and Computer Science (VANC)
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Number of pages
- 66
- Identifiers
- 99901221150501842
- Language
- English
- Resource Type
- Thesis