Active Fire Detection Using a Novel Convolutional Neural Network Based on Himawari-8 Satellite Images
Active Fire Detection Using a Novel Convolutional Neural Network Based on Himawari-8 Satellite Images
复制标题
使用基于 Himawari-8 卫星图像的新型卷积神经网络进行主动火灾检测
DOI：
10.3389/fenvs.2022.794028
复制
发表时间：
2022-03-04
影响因子：
4.6
通讯作者：
Yang, Shuhu
中科院分区：
文献类型：
--
作者：
Hong, Zhonghua;Tang, Zhizhou;Yang, Shuhu
火是一种重要的生态系统过程，在陆地生态系统和大气环境中起着复杂的作用。有时，野火是极具破坏性的自然灾害。为了减少其破坏性影响，必须尽快发现野火。然而，由于传统的阈值方法容易遭受森林小面积砍伐造成的误报，以及浓烟遮蔽的大型火灾的漏报，准确及时地监测野火是一项具有挑战性的任务。深度学习具有学习能力强、适应性强、可移植性好等特点。目前，很少有研究采用深度学习方法以接近真实的时间的方式解决遥感图像中的野火检测问题。因此，在这项研究中，我们提出了一种使用新型卷积神经网络（FireCNN）的主动火灾探测系统。FireCNN采用多尺度卷积和残差接受设计，可以有效提取火点的准确特征。该方法在包含1,823个火点和3,646个非火点的数据集上进行了测试。实验结果表明，FireCNN完全能够进行野火检测，准确率比传统阈值方法高出35.2%。我们还研究了不同的结构设计对神经网络模型性能的影响。对比结果表明，该方法具有较好的效果。
Fire is an important ecosystem process and has played a complex role in terrestrial ecosystems and the atmosphere environment. Sometimes, wildfires are highly destructive natural disasters. To reduce their destructive impact, wildfires must be detected as soon as possible. However, accurate and timely monitoring of wildfires is a challenging task due to the traditional threshold methods easily be suffered to the false alarms caused by small forest clearings, and the omission error of large fires obscured by thick smoke. Deep learning has the characteristics of strong learning ability, strong adaptability and good portability. At present, few studies have addressed the wildfires detection problem in remote sensing images using deep learning method in a nearly real time way. Therefore, in this research we proposed an active fire detection system using a novel convolutional neural network (FireCNN). FireCNN uses multi-scale convolution and residual acceptance design, which can effectively extract the accurate characteristics of fire spots. The proposed method was tested on dataset which contained 1,823 fire spots and 3,646 non-fire spots. The experimental results demonstrate that the FireCNN is fully capable of wildfire detection, with the accuracy of 35.2% higher than the traditional threshold method. We also examined the influence of different structural designs on the performance of neural network models. The comparison results indicates the proposed method produced the best results.