融合CTGAN与机器学习的温州市台风灾害损失评估方法研究OA

Accurate assessment of typhoon-induced disaster losses is often hindered by limited historical data and severe class imbalances,especially in regions with infrequent but high-impact events.These challenges reduce the robustness and generalizability of predictive models,leading to unreliable assessments of potential disaster severity.To address these issues,this study proposes an integrated evaluation method that combines data augmentation using a CTGAN with multiple machine learning algorithms.The objective was to enhance sample diversity,alleviate class imbalance,and improve the accuracy and stability of disaster loss predictions.Wenzhou City,located in Zhejiang Province,China,was selected as the study area because of its frequent exposure to typhoon-related hazards.Twenty typhoon cases from 1994 to 2020 were collected,and a structured dataset was constructed using 13 key indicators.These indicators cover three dimensions:(1)hazard-inducing factors such as maximum wind speed and accumulated rainfall;(2)environmental background conditions,including elevation,river network density,and landform;and(3)socioeconomic exposure and vulnerability,reflected by variables such as population density,GDP per capita,and infrastructure indicators,such as road length and hospital bed count.To represent the level of disaster impact for each event quantitatively,a disaster loss index was calculated and used as the input for k-means clustering.This unsupervised learning approach classified 20 typhoon events into four distinct loss severity levels,forming the basis for subsequent supervised classification tasks.To overcome the limitations of class imbalance,the CTGAN model was employed to generate synthetic samples under specific class-conditional constraints.The generated samples were incorporated into the training set to enrich underrepresented classes and improve the representativeness of the dataset.Five widely used machine-learning models were trained and evaluated:GBDT,XGBoost,LightGBM,CatBoost,and Random Forest.The experimental results demonstrated that the GBDT model outperformed the others in terms of both classification accuracy and generalization performance.This model showed the most consistent results across multiple metrics,including mAP,precision,recall,F1-score,and accuracy.Additionally,a comparative analysis was conducted to explore the influence of synthetic data volume on model performance.The findings revealed that simply increasing the number of synthetic samples does not guarantee continuous improvement;rather,an optimal range of sample sizes exists beyond which model stability may plateau or even decline.This study provides a practical and scalable methodological framework for typhoon disaster loss assessments in data-constrained environments.By leveraging generative modeling and ensemble learning techniques,this study offers insights into the effective application of data-driven methods to support disaster preparedness,emergency response planning,and resilience analysis in other hazard-prone regions.