News thumbnail
Science / Tue, 17 Nov 2026 Nature

Risk assessment and strategic decision support to enhance preparedness and minimize human and economic losses in managing natural disasters

Accurate prediction of the impacts of natural disasters is essential for effective risk management, emergency response, and policy planning. However, the complex, nonlinear, and highly irregular nature of disaster-related data poses significant challenges for traditional modeling approaches. This study proposes a comprehensive machine learning and deep learning framework for economic loss prediction, integrating advanced preprocessing, optimization, and validation strategies to enhance predictive accuracy and robustness. The results demonstrate that the proposed framework achieves robust, statistically significant improvements in predictive performance, highlighting its suitability for reliable economic loss modeling. This study contributes a validated, generalizable, and decision-oriented modeling approach that can support data-driven disaster risk reduction and management strategies.

Accurate prediction of the impacts of natural disasters is essential for effective risk management, emergency response, and policy planning. However, the complex, nonlinear, and highly irregular nature of disaster-related data poses significant challenges for traditional modeling approaches. This study proposes a comprehensive machine learning and deep learning framework for economic loss prediction, integrating advanced preprocessing, optimization, and validation strategies to enhance predictive accuracy and robustness. Multiple predictive models, including deep neural networks, convolutional neural networks, and gradient boosting, are systematically evaluated and optimized using a Turbulent Flow of Water-based Optimization algorithm. To address data irregularities, Isolation Forest is used for outlier detection, and statistical tests are applied to assess distributional assumptions before normalization. Model performance is rigorously assessed using 5-fold cross-validation and a diverse set of evaluation metrics that capture accuracy, bias, and agreement. In addition, ensemble learning is employed to improve predictive reliability further, while runtime analysis provides insights into computational efficiency. The statistical significance of performance differences is verified using ANOVA and the Wilcoxon test, and an ablation study is conducted to examine model stability and sensitivity. The results demonstrate that the proposed framework achieves robust, statistically significant improvements in predictive performance, highlighting its suitability for reliable economic loss modeling. This study contributes a validated, generalizable, and decision-oriented modeling approach that can support data-driven disaster risk reduction and management strategies.

© All Rights Reserved.