Water saturation is a critical parameter determining reservoir fluid properties and productivity, and its accurate assessment is of paramount importance. Current methods for predicting water saturation based on seismic data exhibit strong dependence on petrophysical models, rendering them unsuitable for complex reservoirs. Traditional neural network-based approaches, while independent of petrophysical models, require substantial sample data and struggle to yield satisfactory results in areas with limited well data. To address this issue, we propose a water saturation prediction method based on Generative Adversarial Networks (GAN) built upon precise Zoeppritz inversion using particle filtering. This method integrates Convolutional Neural Networks (CNN) and BP neural networks to design an adversarial neural network suitable for small sample data. Application in the AY gas field demonstrates that, compared to conventional neural networks, the adversarial neural network significantly enhances prediction accuracy for water saturation with limited samples, providing reliable parameters for reservoir evaluation.