关键词: 地表可燃物含水率, 预测模型, 气象要素回归法, 卷积神经网络

Abstract: 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 R² 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 R², 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.

Key words: water content of surface fuel, forecast model, meteorological factors regression, convolutional neural network