Scholarlink

<p>The escalating threat of wildfires to global ecosystems, human life, and property underscores the urgent need for accurate prediction of their spread. The Los Angeles fires of January 2025 serve as a stark reminder of this imperative. Consequently, accurately predicting the scale of wildfire spread and identifying the pivotal factors driving it are essential for formulating effective fire management strategies and emergency response protocols. The limitations inherent in existing research, particularly those pertaining to data imbalance and model interpretability, significantly undermine the credibility of model predictions and their practical applicability. To address these issues, we propose EWSHS, an interpretable wildfire spread scale prediction model based on the Histogram Gradient Boosting (HGB) algorithm. Using historical wildfire data from Alberta, Canada, we first preprocessed the dataset through missing value imputation, standardization, outlier treatment, and categorical feature encoding to enhance feature representation. Subsequently, the Tomek Links undersampling method is applied to address class imbalance, establishing a robust foundation for model training. Comparative analysis reveals that the EWSHS model achieves an accuracy of 91.87%, precision of 91.14%, F1-score of 91.33%, AUC of 0.984, and Matthews Correlation Coefficient (MCC) of 83.19%, achieving better performance than baseline models across all metrics. Finally, integrating the SHAP framework enhances model interpretability by elucidating the key drivers of wildfire spread scale, such as air humidity, wind speed, spread rate, and firefighting interventions. This transparency provides decision-makers with a clear understanding of the model's underlying logic. Specifically, SHAP analysis reveals that the scale of wildfire spread increases dramatically when wind speed exceeds 4&#xa0;m/s and humidity drops below 25%. By transforming numerical outputs into actionable, evidence-based insights, this interpretable framework significantly enhances the reliability and practical value of wildfire predictions for emergency response.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing wildfire spread scale prediction with explainable AI: a hist gradient boosting and SHAP-based approach

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Bin Liao,
  • Tao Zhou,
  • Jiong Yu,
  • Yanping Liu,
  • Tao Zhang,
  • Tiantian Lu

摘要

The escalating threat of wildfires to global ecosystems, human life, and property underscores the urgent need for accurate prediction of their spread. The Los Angeles fires of January 2025 serve as a stark reminder of this imperative. Consequently, accurately predicting the scale of wildfire spread and identifying the pivotal factors driving it are essential for formulating effective fire management strategies and emergency response protocols. The limitations inherent in existing research, particularly those pertaining to data imbalance and model interpretability, significantly undermine the credibility of model predictions and their practical applicability. To address these issues, we propose EWSHS, an interpretable wildfire spread scale prediction model based on the Histogram Gradient Boosting (HGB) algorithm. Using historical wildfire data from Alberta, Canada, we first preprocessed the dataset through missing value imputation, standardization, outlier treatment, and categorical feature encoding to enhance feature representation. Subsequently, the Tomek Links undersampling method is applied to address class imbalance, establishing a robust foundation for model training. Comparative analysis reveals that the EWSHS model achieves an accuracy of 91.87%, precision of 91.14%, F1-score of 91.33%, AUC of 0.984, and Matthews Correlation Coefficient (MCC) of 83.19%, achieving better performance than baseline models across all metrics. Finally, integrating the SHAP framework enhances model interpretability by elucidating the key drivers of wildfire spread scale, such as air humidity, wind speed, spread rate, and firefighting interventions. This transparency provides decision-makers with a clear understanding of the model's underlying logic. Specifically, SHAP analysis reveals that the scale of wildfire spread increases dramatically when wind speed exceeds 4 m/s and humidity drops below 25%. By transforming numerical outputs into actionable, evidence-based insights, this interpretable framework significantly enhances the reliability and practical value of wildfire predictions for emergency response.