Earth fissures caused by tectonic forces, human activities, or both seriously threaten the safety of people’s lives and properties. The Taiyuan Basin, a Cenozoic downfaulted basin located in the centre of the Fen-Wei Basin tectonic belt, in northwestern China, presents the ideal study area for a hazard assessment of earth fissures. A total of 104 earth fissures have been observed in the Taiyuan Basin, with a total length of approximately 128 km. In this paper, we proposed a probabilistic method for mapping earth fissure hazards by integrating the analytic hierarchy process (AHP), the area under the curve (AUC), and the certainty factor model (CFM). Geomorphic units, geologic formations, active faults and land subsidence zones of the Taiyuan Basin were mapped in detail. Correlations between these factors and earth fissures were evaluated through spatial modelling in ArcGIS. The AUC was introduced into the AHP to weight each factor and thus, to derive an earth fissure susceptibility map. Finally, the modelled earth fissure susceptibility was compared with a digital inventory of earth fissures to develop a probability function and map the spatial variability in failure probability through the CFM. The study indicates that active faults have the greatest contribution to the generation of earth fissures. Earth fissures are prone to develop in the piedmont alluvial-diluvial clinoplain and the transitional zone near the geomorphic boundary. This mapping procedure can assist in making rational decisions regarding urban planning and infrastructure development in areas susceptible to earth fissures.

The use of deep learning approaches to predict landslide displacement based on monitored time-series data is an effective method for the early-warning of landslides. Currently, most prediction models focus on the temporal correlation of displacements from a single monitoring point, ignoring the spatial influence of other monitoring points. To fully consider the spatiotemporal features of the displacement data, this paper develops three deep learning models based on graph convolution networks to spatiotemporally predict the landslide displacements of the Huanglianshu landslide. Specifically, we first establish a fully connected graph to represent the spatial relationships of all the deployed monitoring points. Second, we develop a temporal graph convolutional network-long short term memory (TGCN-LSTM) model and an Attention-TGCN model based on the temporal graph convolutional network-gate recurrent unit (TGCN-GRU) deep learning model and employ the three models to spatiotemporally predict displacements of the Huanglianshu landslide. The proposed spatiotemporal prediction models accurately predict the displacements at seven monitoring points, with a maximum R² of 0.85 at the individual monitoring points. The comparative results show that the proposed Attention-TGCN model achieves the highest spatiotemporal prediction accuracy, and the accuracy of the Attention-TGCN model can further improve after considering the movement of the monitoring points.