A neural network model integrating multi-source information was proposed to address the challenge of spatiotemporal feature coupling modeling for embankment settlement prediction in soft soil areas with complex stratigraphy. Stratigraphic profile maps were used as the model input. A convolutional neural network (CNN) was employed to extract the spatial distribution characteristics of soil layers. The feature extraction was enhanced with residual connections and a self-attention mechanism. Simultaneously, groundwater level and historical settlement data were combined to form a bivariate time series. This series was modeled using a bidirectional long short-term memory network (BiLSTM). The optimal parameter combination was identified through the Bayesian optimization algorithm. Based on training, validation, and testing results from the measured dataset of the Xixian Datang site, the model achieves a mean absolute error, root mean squared error, and coefficient of determination of 0.028 6 mm, 0.04 mm, and 0.956 9, respectively. These results demonstrate the superiority of the proposed model over comparison models and confirm its effectiveness and engineering application value for surface settlement prediction under complex geological conditions.