Scholarlink

<p>Spatial modeling of wildfire probability is a key preventive measure for reducing the ecological, economic, and social impacts of wildfires. This study develops an advanced wildfire susceptibility framework for the UNESCO Global Geopark Djerdap (Serbia) by integrating geographic information systems (GIS), multi-sensor satellite data, and state-of-the-art artificial intelligence (AI) algorithms. Four modeling approaches were evaluated: convolutional neural networks (CNN), deep neural networks (DNN), Kolmogorov–Arnold networks (KANs), and Extreme Gradient Boosting (XGBoost). The dataset comprised 1,354 historical wildfire incidents (2001–2024) derived from MODIS and VIIRS imagery, combined with 15 predictive variables representing topographic, climatological, hydrological, vegetational, and anthropogenic factors. The results indicated that approximately 96.79&#xa0;km² of the geopark exhibits high wildfire susceptibility. Among the tested models, XGBoost achieved the highest accuracy (99.38%) and specificity (99.63%), while CNN and KANs demonstrated balanced and robust predictive performance; DNN showed comparatively limited generalization. SHAP (SHapley Additive exPlanations) analysis identified distance from settlements and consecutive dry days as dominant predictors across models. This study presents the first application of Kolmogorov–Arnold Networks in wildfire susceptibility assessment and demonstrates their integration with conventional AI models for explainable spatial prediction. The fusion of multi-sensor data and explainable AI enhances model transparency, supporting the identification of vulnerable settlements and cultural heritage areas. The proposed framework contributes to scalable and interpretable wildfire risk modeling, offering valuable guidance for sustainable landscape management and informed policy-making across protected UNESCO sites. </p> Graphical Abstract <p></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

Machine Learning and Deep Learning Approaches for Wildfire Susceptibility Prediction: A Case Study of the Djerdap Geopark, Serbia

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Uroš Durlević,
  • Nina Čegar,
  • Velibor Ilić,
  • Aleksandar Kovjanić

摘要

Spatial modeling of wildfire probability is a key preventive measure for reducing the ecological, economic, and social impacts of wildfires. This study develops an advanced wildfire susceptibility framework for the UNESCO Global Geopark Djerdap (Serbia) by integrating geographic information systems (GIS), multi-sensor satellite data, and state-of-the-art artificial intelligence (AI) algorithms. Four modeling approaches were evaluated: convolutional neural networks (CNN), deep neural networks (DNN), Kolmogorov–Arnold networks (KANs), and Extreme Gradient Boosting (XGBoost). The dataset comprised 1,354 historical wildfire incidents (2001–2024) derived from MODIS and VIIRS imagery, combined with 15 predictive variables representing topographic, climatological, hydrological, vegetational, and anthropogenic factors. The results indicated that approximately 96.79 km² of the geopark exhibits high wildfire susceptibility. Among the tested models, XGBoost achieved the highest accuracy (99.38%) and specificity (99.63%), while CNN and KANs demonstrated balanced and robust predictive performance; DNN showed comparatively limited generalization. SHAP (SHapley Additive exPlanations) analysis identified distance from settlements and consecutive dry days as dominant predictors across models. This study presents the first application of Kolmogorov–Arnold Networks in wildfire susceptibility assessment and demonstrates their integration with conventional AI models for explainable spatial prediction. The fusion of multi-sensor data and explainable AI enhances model transparency, supporting the identification of vulnerable settlements and cultural heritage areas. The proposed framework contributes to scalable and interpretable wildfire risk modeling, offering valuable guidance for sustainable landscape management and informed policy-making across protected UNESCO sites.

Graphical Abstract