TY - JOUR
T1 - Real-time verification of solar-powered forest fire detection system using ensemble learning
AU - Yıldıran, Nezihe
N1 - Publisher Copyright:
© 2024 Elsevier Ltd
PY - 2024/12/1
Y1 - 2024/12/1
N2 - In recent years, the surge in forest fires has led to widespread devastation, posing significant threats to both the environment and human populations. Timely detection of fires is crucial in containing their spread, and recent advancements in deep learning hold promise for proactive detection of such incidents. This research paper presents a comprehensive analysis of a substantial dataset comprising 25,015 images encompassing fire, smoke, non-fire, day, and night. The dataset was utilized for pre-training with fifteen models, including VGG16, ResNet50V2, DenseNet201, and ConvNeXtLarge. The most effective models were integrated into ensemble learning, yielding a validation accuracy of 96.54% and a testing accuracy of 92.04% using weighted majority ensemble learning. The algorithm was meticulously developed using the parallel computing infrastructure of the TRUBA high-performance computing center, employing 120 servers, 20 cores x 2 CPU, resulting in expedited execution. Furthermore, a real-time prototype of the fire detection system was implemented to validate the practical applicability of the proposed algorithm. This work not only introduces substantial innovation in fire detection methodologies but also signifies a significant stride towards the establishment of an efficient early warning system for forest fires. The study relies in its comprehensive analysis of a large, diverse dataset, the integration of multiple deep learning models, and the real-time validation of the developed algorithm, all of which contribute to the advancement of fire detection research.
AB - In recent years, the surge in forest fires has led to widespread devastation, posing significant threats to both the environment and human populations. Timely detection of fires is crucial in containing their spread, and recent advancements in deep learning hold promise for proactive detection of such incidents. This research paper presents a comprehensive analysis of a substantial dataset comprising 25,015 images encompassing fire, smoke, non-fire, day, and night. The dataset was utilized for pre-training with fifteen models, including VGG16, ResNet50V2, DenseNet201, and ConvNeXtLarge. The most effective models were integrated into ensemble learning, yielding a validation accuracy of 96.54% and a testing accuracy of 92.04% using weighted majority ensemble learning. The algorithm was meticulously developed using the parallel computing infrastructure of the TRUBA high-performance computing center, employing 120 servers, 20 cores x 2 CPU, resulting in expedited execution. Furthermore, a real-time prototype of the fire detection system was implemented to validate the practical applicability of the proposed algorithm. This work not only introduces substantial innovation in fire detection methodologies but also signifies a significant stride towards the establishment of an efficient early warning system for forest fires. The study relies in its comprehensive analysis of a large, diverse dataset, the integration of multiple deep learning models, and the real-time validation of the developed algorithm, all of which contribute to the advancement of fire detection research.
KW - Deep learning
KW - Early warning system
KW - Fire detection
KW - Forest fire
KW - Weighted majority
UR - http://www.scopus.com/inward/record.url?scp=85199404260&partnerID=8YFLogxK
U2 - 10.1016/j.eswa.2024.124791
DO - 10.1016/j.eswa.2024.124791
M3 - Article
AN - SCOPUS:85199404260
SN - 0957-4174
VL - 255
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 124791
ER -