Abstract: Fires are a ubiquitous and frequent occurrence all over the world.However，urban fire risk prediction is still in its infancy.Therefore，this study proposes a prediction model HISS for urban fire risk.Visual features from satellite images are extracted based on multi-fractal dimensions and fused with non-visual features extracted from historical data on points of interest such as fire，meteorology，inflow and outflow of cabs，and electricity usage to form static features.The static features are pre-trained using XGBoost，and a feature embedding module is proposed.The prediction results of XGBoost are embedded as a benchmark into the dynamic features with temporal relations，whereby the fluctuation patterns of fire are learned through a transformer.A dynamic weighting module is proposed to fuse XGBoost and the transformer at the model level to further improve model performance.The experimental results demonstrate that the determination coefficient R² of the proposed model reached 72.56%，and the performance of the model compared with the Long Short-Term Memory（LSTM） time series model and Gated Recurrent Unit（GRU） improved by 4.25 and 3.92 percentage points，respectively.Compared with Lasso，Random Forest（RF），and Gradient Boosting Decision Tree （GBDT），the performance of the proposed model is improved by 10.88，5.62，and 3.93 percentage points，respectively.The HISS model has excellent predictive performance.

Key words: spatial-temporal big data, fire risk prediction, satellite images, feature embedding, dynamic weighting

摘要： 火灾事故在全世界范围内普遍存在且频繁发生。然而，针对城市火灾的风险预测研究尚在起步中。为此，提出一种针对城市火灾风险的预测模型HISS。基于多重分形维度从卫星图像中提取视觉特征，并将卫星图像的视觉特征与从历史火灾数据、气象数据、出租车流动记录、区域用电量数据和POI数据中提取的非视觉特征相融合，构成静态特征。基于静态特征对XGBoost进行预训练，并设计一种特征嵌入模块，将XGBoost的预测结果作为基准值嵌入到包含时序关系的动态特征中，通过Transformer学习火灾的时间波动模式。采用动态加权模块实现XGBoost和Transformer在模型层面的融合，进一步提升模型的性能。实验结果表明，HISS模型的确定系数R²达到了72.56%，与长短期记忆网络和门控循环单元相比R²分别提升4.25和3.92个百分点，与Lasso、随机森林和梯度提升决策树相比R²分别提升10.88、5.62和3.93个百分点，具有较优的预测性能。

关键词: 时空大数据, 火灾风险预测, 卫星图像, 特征嵌入, 动态加权

CLC Number: 

• TP391