With the acceleration of global urbanization, urban expansion prediction faces the problems of insufficient fusion of multi-source data and inaccurate capture of spatiotemporal dynamic features, which leads to the accuracy bottleneck of existing models in modeling complex nonlinear relationships. This study proposes a deep learning model that integrates multi-source spatiotemporal data to improve the accuracy and spatial expression ability of urban expansion prediction. The model uses ConvLSTM as the core architecture, integrates Landsat remote sensing images, POI points of interest, traffic road network and terrain data from 1990 to 2020, constructs a spatiotemporal cube through data cleaning, normalization and spatial rasterization (resolution 30 m × 30 m), and uses a sliding window to generate sequence samples; the network contains three layers of ConvLSTM units (64-128-256 channels) to extract spatiotemporal features, embeds a spatial attention mechanism to strengthen the identification of construction hotspots, and finally outputs a probability map through a fully connected layer. The experiment uses a five-fold cross-validation, and the test in the Yangtze River Delta urban agglomeration shows that the overall accuracy of the model is as high as 94.3%, the average Kappa coefficient is 0.85, and the mean square error of construction land is reduced to 0.086. The results show that the model can effectively explore the spatiotemporal correlation patterns of urban expansion, verify the significant advantages of deep learning in urban simulation, and provide a high-precision decision-making tool for national land space planning. In the future, the model’s lightweight deployment and real-time prediction functions can be expanded.