Landslide spatial prediction is often hindered by conventional models that conflate long-term predisposing factors with short-term triggering signals, limiting their ability to capture, disentangle, and integrate spatiotemporal information. To address this limitation, we propose a new method that explicitly distinguishes spatial constraints and temporal trends. Specifically, a graph neural network guided by the Third Law of Geography is used to learn representations from stable spatial features, while a temporal Transformer encodes the timevarying dynamics of surface deformation. The resulting spatial and temporal embeddings are then fused to enable effective cross-modal information aggregation. We implement and validate the proposed framework in Luding County and the upper Jinsha River region. Results show that, across three mapping units, our method consistently outperforms representative machine-learning baselines (e.g., spatial-only and temporal-only strategies), yielding the improvements of 8.7%, 13.0%, and 8.3% in the area under the receiver operating characteristics curve, respectively. Moreover, slope units provide better predictive performance than grid and triangular irregular network (TIN) units for landslide susceptibility modelling. To our knowledge, this is the first approach to explicitly model spatial and temporal characteristics as heterogeneous inputs, treating long-term predisposing factors and InSAR-derived deformation sequences as distinct modalities, thereby addressing a key challenge in multi-modal feature extraction. These findings suggest that the proposed framework offers a promising solution for landslide spatial prediction and could support operational early warning when integrated with external triggering indicators.