You Only Look Twice: An Ensemble Deep Learning Model for Wildfire Detection Using Terrestrial Camera Networks

Wildfires represent a growing global threat that requires rapid detection and response to minimize environmental damage, economic losses, and human casualties. In the United States, California stands out as a particularly common wildfire hot spot. Recent fire seasons have shattered historical records and been particularly devastating. This work investigates innovative methods for classifying and localizing wildfires through terrestrial cameras positioned on elevated terrain, aimed at improving early detection capabilities and response times while maintaining computational efficiency and reliability for the U.S. Space Force in Southern California. We present YOL2, a novel ensemble approach that combines a fine-tuned ConvNeXt Convolutional Neural Network incorporating a Dynamic Tanh normalization layer with a fine-tuned YOLO11 model for precise localization. Using a comprehensive dataset of 33,636 time-sequenced images from terrestrial cameras across the United States and Europe, our system achieves 98% fire detection accuracy and 55% localization mean average precision [50:95]. The implementation of Dynamic Tanh normalization—applied for the first time in wildfire detection—enhances computational efficiency without sacrificing performance. The images used capture the spread of incipient fires over time, with most containing bounding boxes denoting the approximate location of fire, allowing our system to identify fires quickly while minimizing false positives. Importantly, our spatiotemporal system operates effectively without requiring individual models to rely on multiple time steps as input, enabling modular component replacement and adaptation. The use of pan, tilt, and zoom cameras in concert with our YOLO model provides a more computationally efficient confirmation of fire than alternative methods, showing that extracting better results from less information is possible. Beyond wildfire applications, the YOL2 ensemble methodology demonstrates profound implications for remote sensing more broadly. This work establishes a foundation for highly efficient visual detection systems applicable across numerous domains requiring rapid and accurate object identification and localization.

Name thesis:degree_name

Master

Department dc:contributor.department

Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science

Grantor dc:publisher

Massachusetts Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Jones, John M.

Advisor dc:contributor.advisor

• Murray, Fiona

Statement dc:rights

• In Copyright - Educational Use Permitted
• Copyright retained by author(s)

Licence dc:rights.uri

• https://rightsstatements.org/page/InC-EDU/1.0/

Handle dc:identifier.uri

https://hdl.handle.net/1721.1/162731

OAI identifier oai:identifier

oai:dspace.mit.edu:1721.1/162731