Research on a Small-Sample Post-Earthquake Population Prediction Method Based on Improved WGAN-Based Sample Augmentation

Accurately estimating the number of individuals impacted by earthquakes is essential for prompt response and effective allocation of emergency resources. Conventional models tend to overfit and exhibit poor generalization, particularly when dealing with limited datasets devoid of historical earthquake data. In this research, we introduce a WGAN-GP-T technique leveraging generative adversarial networks to produce synthetic samples for forecasting impact figures in scenarios with scarce data. This method incorporates a filter to enhance the quality of generated data and facilitate model training. To address the multidimensional characteristics of earthquake data, we propose a quality assessment strategy utilizing an attribute-weighted average sample overlap rate to sift through and refine synthetic samples. Empirical findings demonstrate that utilizing the generated synthetic samples decreases the root mean square error (RMSE) of predictive models by an average of 69.05%. Following the screening process, models necessitate fewer supplementary samples and achieve further RMSE reduction, yielding an average enhancement of 27.17% under constrained sample conditions. The proposed approach effectively supplements original datasets, thereby enhancing modeling efficacy and generalization performance for post-earthquake impact forecasting.