Abstract. In recent years, the frequency of landslide disasters has been increasing year by year due to the extension of human activities to the natural environment. Fast and detailed landslide surveys are important for landslide disaster prediction and management. There are many driving factors for landslide formation, and most of the current deep learning-based landslide identification methods use optical remote sensing images in a short period or a few types of fused data for prediction. Therefore the upper limit of accuracy they can achieve is low. This paper proposes a landslide identification network model based on the spatio-temporal fusion of heterogeneous data from multiple sources. The model takes observations such as time-series optical remote sensing images, DEM, geological formations, and meteorological data as inputs. To address the problems of non-uniform data forms and redundancy caused by time-series data, we design the temporal phase fusion module of coupled CNN-LSTM to fuse the temporal features of multi-source data based on the extraction of their spatial features. Subsequently, we design the spatial feature fusion module based on DCNN-DBN to realize the deep expression of temporal phase and spatial features of landslides and improve the recognition efficiency and accuracy of the network. Through experimental verification, the AUC value of our proposed model is 0.8976, the F1 score is 0.8352, and the MIoU is 0.8624. The evaluation results reflect that the model can provide support for large-scale landslide disaster investigation.