基于缓冲区重采样的LSTM林火预测模型*

为提高林火风险预测精度，挖掘地图上隐含的空间信息、时间序列上隐含的长期趋势和循环波动，提出1种基于缓冲区重采样的长短期记忆（LSTM）林火预测模型，选取15个与林火相关的影响因素，以方差膨胀因子为评价指标对其进行多重共线性检验，方差膨胀因子大于10的因素具有共线性，并采用信息增益率验证筛选结果的合理性。考虑到火灾的空间聚集特性，采用缓冲区分析与过采样相结合方法减少样本不均衡现象的影响，最终得到176 732条样本。对12个影响因素和研究时间段的火点建立LSTM预测模型，对森林火灾发生风险进行预测。研究结果表明:基于缓冲区重采样的LSTM林火预测模型有效考虑时空上隐含的信息，预测模型准确率为87.06%，特异性为97.99%，敏感度为76.12%，阳性预测率为97.43%，阴性预测率为80.41%，ROC曲线与AUC值均优于随机森林（RF）和支持向量机（SVM）这2种基准算法。维尔克松秩和检验发现，本文提出的模型与基准算法结果具有显著性差异。研究结果可为提高林火风险预测精度提供参考。

LSTM prediction model of forest fire based on buffer resampling

In order to improve the prediction accuracy of forest fire risk,the spatial information hidden in the map and the long-term trend and cyclic fluctuation hidden in the time series were mined.A long-short term memory (LSTM) prediction model of forest fire based on the buffer resampling was proposed,then 15 influencing factors related to forest fire were selected,and the variance expansion factor was used as the evaluation index to carry out the multicollinearity test.The factors with variance expansion factor greater than 10 had the collinearity,and the information gain rate was used to verify the rationality of the screening results.Considering the spatial aggregation characteristics of fire,the combination method of buffer analysis and oversampling was used to reduce the impact of sample imbalance,and 176 732 samples were finally obtained.The LSTM prediction model for 12 influencing factors and fire points in the research period was established to predict the risk of forest fire.The results showed that the LSTM prediction model of forest fire based on buffer resampling effectively considered the hidden information in time and space.The accuracy of the prediction model was 87.06%,the specificity was 97.99%,the sensitivity was 76.12%,the positive prediction rate was 97.43%,and the negative prediction rate was 80.41%.Both the ROC curve and AUC value were better than the two benchmark algorithms of random forest (RF) and support vector machine (SVM).The Wilcoxon rank sum test indicated that the results of the model proposed in this study were significantly different from those of the benchmark algorithms.The research results can provide reference for improving the prediction accuracy of forest fire risk.