基于山火时空特性的深度学习火点识别模型
Deep Learning Fire Spot Identification Model Based on Temporal and Spatial Characteristics of Mountain Fire
摘要: 山火风险评估对火灾管理有着重要的作用。本文将NASA-FIRM网站上获取的2018~2019年的山火火点以及电网巡检发现的火点作为实验数据集，通过调研山火的影响因子，最终从天气数据、遥感数据、地形因素、人为因子四个方面选取了降雨量、平均相对湿度、最高气温、最大阵风风速、最大阵风风向、植被归一化指数、植被含水率、海拔、坡度、坡向、距离道路的距离、距离河流的距离、距离村庄的距离15个因素作为山火的影响因子。我们将数据集输入到ConvLSTM模型分类器进行训练验证。基于不同山火影响因子对山火的影响程度不同的思想，我们提出在ConvLSTM上添加通道注意力机制，让模型在训练的过程中学习各个通道对山火影响的权重，对ConvLSTM添加通道注意力机制(Attention ConvLSTM, Att-ConvLSTM)，对模型重新进行训练，实验结果显示，改进后的通道注意力ConvLSTM模型在验证集对山火的识别准确率提升了5.7%，改进后的模型预测和真实情况更加吻合。这使得模型能够更准确地根据山火影响因子对输电走廊附近区域进行山火风险评估预测，使电网单位能够及时发布预警消息。
Abstract: Mountain fire risk assessment plays an important role in fire management. In this paper, mountain fires in 2018~2019 obtained from NASA-FIRM website and those found by power grid inspection are used as experimental data sets. Through the investigation of the influencing factors of mountain fire, 15 factors, such as rainfall, average relative humidity, maximum temperature, maximum gust wind speed, maximum gust direction, vegetation normalization index, vegetation moisture content, altitude, slope, slope direction, distance from road, distance from river and distance from village, are selected as influencing factors from four aspects: weather data, remote sensing data, topographic factors and human factors. We input the data set to ConvLSTM model classifier for training verification. Based on the idea that different mountain fire influence factors have different degrees of influence on mountain fire, we proposed to add channel attention mechanism to ConvLSTM to make the model learn the weight of influence of each channel on mountain fire during training. Attention ConvLSTM (ATT-ConvLSTM) was added to the ConvLSTM, and the model was retrained. Ex-perimental results showed that the improved ConvLSTM improved the accuracy of mountain fire identification by 5.7% in the verification set. The improved model’s predictions are in better agreement with the real world. This enables the model to more accurately assess and predict the risk of mountain fire in the vicinity of the transmission corridor according to the impact factors of mountain fire, and enables the power grid units to issue early warning messages in time.
文章引用：刘星星, 邓杰航, 徐国涛, 刘栋濠, 杨析睿, 陈树东, 冯子垚. 基于山火时空特性的深度学习火点识别模型[J]. 计算机科学与应用, 2021, 11(9): 2369-2377. https://doi.org/10.12677/CSA.2021.119242

参考文献

[1] 刘春翔, 范鹏, 王海涛, 郭江, 柯睿. 基于BP神经网络的输电线路山火风险评估模型[J]. 电力系统保护与控制, 2017, 45(17): 100-105.
[2] 文斌, 谢献强, 孙萌, 杜治国, 李溯, 黄平, 朱宇浩, 谢柏联. 基于加权Logistic回归模型的森林火灾预测[J]. 林业与环境科学, 2019, 35(4): 79-83.
[3] 郑焕能, 居恩德. 林火管理[M]. 哈尔滨: 东北林业大学出版社, 1988: 15-18.
[4] Vadrevu, K.P., Eaturu, A. and Badarinath, K.V.S. (2010) Fire Risk Evaluation Using Multicriteria Analysis—A Case Study. Environmental Monitoring and Assessment, 166, 223-239. [Google Scholar] [CrossRef] [PubMed]
[5] Lee, S. and Pradhan, B. (2007) Landslide Hazard Mapping at Se-langor, Malaysia Using Frequency Ratio and Logistic Regression Models. Landslides, 4, 33-41. [Google Scholar] [CrossRef] 
[6] Shi, S., Yao, C., Wang, S., et al. (2018) A Model Design for Risk Assessment of Line Tripping Caused by Wildfires. Sensors, 18, 1941. [Google Scholar] [CrossRef] [PubMed]
[7] Pham, B.T., Jaafari, A., Avand, M., et al. (2020) Performance Evaluation of Machine Learning Methods for Forest Fire Modeling and Prediction. Symmetry, 12, 1022. [Google Scholar] [CrossRef] 
[8] Zhang, G., Wang, M. and Liu, K. (2019) Forest Fire Susceptibility Modeling Using a Convolutional Neural Network for Yunnan Province of China. International Journal of Disaster Risk Science, 10, 386-403. [Google Scholar] [CrossRef] 
[9] Hochreiter, S. and Schmidhuber, J. (1997) Long Short-Term Memory. Neural Computation, 9, 1735-1780. [Google Scholar] [CrossRef] [PubMed]
[10] Shi, X., Chen, Z., Wang, H., Yeung, D.-Y., Wong, W.-K. and Woo, W.-C. (2015) Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. Pro-ceedings of the 28th International Conference on Neural Information Processing Systems, 1, 802-810.
[11] 覃先林. 遥感与地理信息系统技术相结合的林火预警方法的研究[D]: [博士学位论文]. 北京: 中国林业科学研究院, 2005.