基于卷积神经网络及气象要素回归法的地表死可燃物含水率预测模型

Prediction model of water content in surface dead fuel based on convolution neural network and meteorological factors regression

地表可燃物含水率是森林火险等级和火行为变化的重要指标,其预测模型对于火险预测、火灾管理等具有显著作用。本研究基于蒙古栎及樟子松林地的野外气象因子以及地表死可燃物含水率数据,进行气象因子随机森林相对重要性排序以及皮尔逊相关性分析,并使用深度学习中的卷积神经网络以及气象要素回归法预测可燃物含水率。结果表明:野外蒙古栎的可燃物含水率显著高于樟子松。随机森林结果表明,对于可燃物含水率具有显著作用的因子排列顺序从大到小为湿度、温度、降雨、风速、太阳辐射;相关性分析表明,当日的温度、湿度、降雨对于可燃物含水率具有显著影响,同时,气象因子之间也存在一定的相关性。卷积神经网络模型对于蒙古栎及樟子松林地表可燃物含水率的预测R~2分别为0.928、0.905,平均绝对误差(MAE)分别为6.1%、8.1%,平均相对误差(MRE)分别为8.9%、4.2%;气象要素回归法的R~2分别为0.495、0.525,MAE分别为30.5%、39.5%,MRE分别为52.1%、32.6%,卷积神经网络模型精度显著高于气象要素回归法。研究表明,深度学习的卷积神经网络能够为今后的可燃物含水率预测提供一定借鉴,可为更高水平的林火管理提供有效技术支撑。

Water content of surface fuels is an important indicator of forest fire risk level and fire behavior,and the prediction model of which has a significant effect on fire risk prediction and management.Based on field meteorological factors of Quercus mongolica and Pinus sylvestris var.mongolica forests and water content data of dead fuels on the ground,we conducted the relative importance ranking of meteorological factors random forest and Pearson correlation analysis,and predicted water content of fuels using deeply learned convolutional neural network and meteorological factors regression.The results showed that water content of Q.mongolica fuels in the wild was significantly higher than that of P.sylvestris var.mongolica.The results of random forest showed that the factors that had significant effect on water content of fuel were humidity,temperature,rainfall,wind speed,and solar radiation,with the importance ranking from the largest to the smallest.Results of correlation analysis showed that temperature,humidity,and rainfall of current day had a significant impact on water content of fuels,and certain correlations were observed between meteorological factors.The prediction R?of the convolutional neural network model for the surface fuel water content of Q.mongolica and P.sylvestris var.mongolica forest was 0.928 and 0.905,the mean absolute error(MAE)was 6.1%and 8.1%,and the mean relative error(MRE)was 8.9%and 4.2%,respectively.However,the R^(2),MAE,MRE of meteorological factors regression were 0.495 and 0.525,30.5%and 39.5%,52.1%and 32.6%,respectively.The precision of convolution neural network model was significantly higher than that of meteorological factors regression.Our results showed that the deeply learned convolutional neural network could provide some reference for the prediction of fuel water content in the future,and effectively support higher level forest fire management.