Page 205 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Cong Tung Dinh, Thu Huong Nguyen, Huyen Do Thi, Nam Anh Bui 189

Fig 6. Confusion matrix of Xception model

As can be seen, all four models are highly accurate in applying forest fire detection
based on images collected from the camera. The ResNet152-V2 model is the most
efficient, with little difference in Precision and Recall ratios, although it incorrectly
predicted 15 cases, ResNet152-V2 only missed 2 cases out of 190. This suggests that
ResNet152-V2 is the model that matches the dataset used in the paper. In the VGG-19
model, the Precision value is the highest, demonstrating the model's effectiveness in
fire detection for predictive data samples. However, if based on total actual data, the
Recall value is lower than the ResNet-152 model. In the Xception model, it may be
because the depth convolution layer is not effective, when the dataset with large fire
images, the surrounding background has the same color and features as fire, creating
confusion, leading to the combination of features on each image channel of Xception
is not highly effective. Therefore, in the Confusion matrix, the number of cases
mistakenly predicted to be fire and not fire is not much different (14 cases and
9 cases). In the Inception-V3 model, although the accuracy is slightly higher than that
of the Xception, the difference between Precision and Recall is large, resulting in an
F1-Score measurement almost on par with the Xception model.

5 Conclusions

Early detection of forest fires is an urgent need today. The paper presented and
evaluated four models of applying deep learning techniques ResNet152-V2, VGG-19,
Inception, and Xception applying forest fire detection through images collected from
cameras. The test results show that the above methods are highly accurate, in which,
the forest fire detection rate of the Xception model is 93.94%, the Inception-V3 model
is 94.21%, the VGG-19 model is 94.73% and the ResNet152V2 model is 95.53%.
The results also show that the fire detection program with the ResNet152-V2 model
has the highest accuracy. In the future, we will develop and improve the deep learning
network model to achieve greater accuracy, which can be effectively applied in forest
fire detection and warning.

ISBN: 978-604-80-8083-9 CITA 2023