Scholarlink

Wildfires pose a significant threat to ecosystems and human life, making early detection critical. Recent advances in deep learning, particularly Convolutional Neural Networks (CNNs), have shown promise in automating wildfire detection using camera and drone imagery. However, these models often operate as “black boxes,” providing little insight into their decision-making processes, which raises concerns about their reliability in real-world applications. This paper tackles this issue by evaluating three CNN models—ResNet-50, Inception-V3, and Xception—on the DeepFire dataset, not only in terms of accuracy but also in interpretability. We apply Grad-CAM, an Explainable AI (XAI) technique, to visualize where each model focuses when predicting fire presence. Our results show that while Xception achieves the highest accuracy, Inception-V3 demonstrates more consistent attention to critical fire-related areas. By integrating Grad-CAM, we enhance model transparency, offering a more reliable and interpretable approach to wildfire detection.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Visual Explanation of Deep Learning Models for Wildfire Detection: A Grad-CAM Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hamza Touati,
  • Said Labed,
  • Hadjir Zemmouri

摘要

Wildfires pose a significant threat to ecosystems and human life, making early detection critical. Recent advances in deep learning, particularly Convolutional Neural Networks (CNNs), have shown promise in automating wildfire detection using camera and drone imagery. However, these models often operate as “black boxes,” providing little insight into their decision-making processes, which raises concerns about their reliability in real-world applications. This paper tackles this issue by evaluating three CNN models—ResNet-50, Inception-V3, and Xception—on the DeepFire dataset, not only in terms of accuracy but also in interpretability. We apply Grad-CAM, an Explainable AI (XAI) technique, to visualize where each model focuses when predicting fire presence. Our results show that while Xception achieves the highest accuracy, Inception-V3 demonstrates more consistent attention to critical fire-related areas. By integrating Grad-CAM, we enhance model transparency, offering a more reliable and interpretable approach to wildfire detection.