【目的】 林火的发生不仅会对森林生态环境造成严重的破坏,影响生态系统功能,还会给人类造成一定危害和损失。基于深度学习对森林火灾图像进行识别,旨在更高效精准地对森林火灾发生初期的图像进行识别并预警,从而降低森林火灾对森林生态系统和人类社会产生的危害。【方法】 借助SSD算法目标检测算法在TensorFlow上的实现,根据林火特征对其网络结构进行适当优化,提出一种可以识别森林火灾图像中火焰特征的模型方法。首先对获取的图像进行归一化处理,然后使用Imgaug图像数据增强库对林火图像进行数据增强以构建林火数据集,搭建深度学习运行环境并设定超参数。通过对测试集中的数据进行测试获取模型对图像中林火特征的识别效果,并利用Loss曲线、P-R曲线的可视化对模型进行评估,最后得到模型对于林火的识别精度。【结果】 随着迭代次数的增加,损失值由训练初期的35.31逐渐下降,训练至20 000步时损失值稳定在7.10左右,此时模型的识别精度达到较高水平,对测试数据中林火特征的识别置信度达到0.9以上。基于FLAME公开数据集中的林火图像搭建测试集,经过测试评估本模型对林火特征的mAP可以达到97.40%,漏检率为0.03,对测试图片的平均检测时间仅为0.07 s,对比Faster R-CNN模型在同等数据集上的实现,SSD可以获得更为理想的检测速度。【结论】 提出的针对林火特征识别的SSD算法能兼顾检测速度和检测精度,对林火早期的图像能够快速识别并拥有较低的漏检率,有助于林区工作人员对火灾及时做出处理,从而为森林火灾早期预防提供技术参考。

【Objective】 Forest fires not only cause serious damage to the natural environment of forests and affect their ecosystem functions, but also cause harm and losses to human society. Based on the use of a deep learning process to identify forest fire images, it is possible to efficiently and accurately identify images showing the early stages of forest fires, thereby enabling prompt action to reduce the harm caused by forest fires to forest ecosystems and human society. 【Method】 With the implementation of the single shot multibox detector (SSD) target detection algorithm based on the TensorFlow framework, the network structure was optimized according to the characteristics of forest fires, and a model was developed that could identify flame features in forest fire images. First, the acquired images were normalized and then the “Imgaug” image data enhancement library was used to perform data enhancement on forest fire images and build a forest fire dataset, followed by the building of a deep learning operating environment and the setting of hyper-parameters. The recognition of forest fire features in the image by the model was obtained by testing the data in the test set. By visualizing both the Loss curve and the precision-recall (P-R) curve, the recognition accuracy of the model for forest fires was ultimately obtained.【Result】 With an increase in the number of iterations, the loss value gradually decreased from 35.31 in the early stage of training, and stabilized at about 7.10 when the training reached 20 000 iterations. At this point, the recognition accuracy of the model reached a high level. The recognition confidence of fire features was greater than 0.9. A test set was built based on the forest fire images in the fire luminosity airborne-based machine learning evaluation (FLAME) public data set. After testing and evaluation, the model achieved a mean average precision for forest fire characteristics of 97.40%, the missed detection rate was 0.03, and the average detection time of the test image was only 0.07 s. Compared with the faster- region-based convolutional neural network (Faster R-CNN) model using the same data set, the SSD achieved a better detection speed and recognition accuracy. 【Conclusion】 The SSD algorithm for forest fire feature recognition proposed here can take into account the detection speed and precision, and can quickly identify images in the early stage of a forest fire with a low missed detection rate, which will enable forest staff to deal with the fire in time, providing a technical reference for the early prevention of forest fires.