Abstract: Aiming at the difficulty of building a data-driven class of site amplification effect prediction model due to the sparse distribution of strong earthquake stations in a local region, this study proposes a method of constructing a local site effect prediction model based on the observation data of multiple seismic events from a few stations. The site amplification index is determined by the Surface/Basement Spectral Ratio (SBSR) and the Generalized Inversion Technique (GIT) for site effect evaluation. It is combined with the Convolutional Neural Networks (CNN) and the Generalized Inversion Technique (GIT). Networks (CNN) and Long Short-term Memory Networks (LSTM), four classes of intelligent prediction models (CNN-SBSR, CNN-GIT, LSTM-SBSRand LSTM-GIT) were constructed. Among them, the SBSR class models selected the source parameters (magnitude, epicenter distance, and epicenter depth), site parameters (shear wave velocity, peak bedrock recorded acceleration), and spatial coordinates (latitude and longitude) as input parameters, while the GIT class models relied on the relevant parameters such as site characteristics and spatial information. The difference with the traditional site prediction method is that the established prediction model can directly learn the observation data of a single seismic event instead of regressing based on the station-averaged spectral ratio curves, which effectively solves the limitation of the traditional method that is difficult to be extrapolated to the station-free region. The cross-validation results of the prediction models constructed based on the proposed method using data from five stations in the Lower Hart Basin, New Zealand, show that the LSTM-like model exhibits higher prediction accuracy by its excellent time-series feature extraction capability, while the CNN model shows better robustness with fewer parameters, and the GIT method outperforms the SBSR method due to the strengthened role of the site parameters. Among them, the combined LSTM-GIT model performs optimally. The technical solution proposed in this study provides a solution with both accuracy and reliability for site amplification prediction in sparse station areas.