基于粒子群算法和BP神经网络的多因素林火等级预测模型

针对现有小尺度林火预测模型预测结果有效性、可扩展性等方面的不足,通过考虑多种火险因素,构建BP神经网络预测模型以提高预测精度,在此基础上借助粒子群算法加快BP神经网络收敛速度,进而提出一种混成的多因素森林火险等级预测模型particle swarm optimization based back-propagation neural network (PSO-BP)。所构建的预测模型,能够同时考虑气候因素(日最高气温、日平均气温、24 h降水量、连旱天数、日照时数、日平均相对湿度、日平均风速)、地形地貌因素(海拔、坡度、坡向、土壤含水量)、可燃物因素(植被类型、可燃物含水率、地被物载量)、人为因素(人口密度、距人类活动区域的距离) 16个变量。基于南京林业大学下蜀林场森林防火实验站传感器网络所采集的实际数据及现场测量数据,通过一组试验验证提出模型的有效性。结果表明:基于训练数据集及检验样本所构建的模型能够开展有效的火险等级预测;模型的计算复杂度较单独使用BP神经网络模型明显下降。

A multi-factor forest fire risk rating prediction model based on particle swarm optimization algorithm and back-propagation neural network

Forest fire destroys woodlands and forest resources, and emits massively greenhouse gases. It has been recognized as a series disaster for the sustainability of forests. Prediction of the risk rating for forest fire early warning has been a hot topic and widely investigated in recent years. To prevent forest fire and mitigate loss, an effective forest fire risk rating prediction desires for a real-time and region-related result. The relevant existing studies have proposed some forest fire risk rating prediction approaches. Nevertheless, the exiting approaches still cannot meet the previously mentioned forest fire prevention application requirements. Firstly, different models fit well for different regions (concerning climates and fuel types, etc.). Seldom approaches can scale for different regions. Secondly, the prediction results are split by days. To perform short-term multi-factor forest fire risk rating prediction and guarantee the accuracy and scalability of prediction model, we investigated the machine learning methodologies in this paper and proposed a particle swarm optimization (PSO)-based back-propagation (PSO-BP) neural network model. Specially, we incorporated the following 16 forest fire dangerous factors based on the existing reports in PSO-BP,i.e., meteorological factors (e.g., daily maximum air temperature, daily average air temperature, 24-hour rainfall, the number of dry days, hours of sunshine, daily average relative humidity, and daily average wind velocity), terrain factors (e.g., altitude, gradient, exposure, and soil water content), combustible factors (e.g., types of vegetation, fuel moisture content, and content of forest litter), and human factor (e.g., density of population, distance to the human activity sites). As for the PSO-BP model, we firstly used BP neural network as the foundation technology to construct the prediction model. To guarantee the performance in terms of computational complexity of the prediction model, we then adopted PSO algorithm to speed up the convergence speed of BP neural network training process. We also presented a case study based on the data collected via the sensor-network-based forest fire prevention experimental station, Xiashu Forest, Nanjing Forestry University, China. The experimental results suggested that: 1) The PSO-BP prediction model constructed based on the training data set can effectively predict the forest fire danger ratings in the near-future for the study region;and 2) The efficiency in terms of computational complexity for training the prediction model using the PSO-BP is better than that of those models using solely BP neural network.