Objectives Aiming at the current geological disaster prevention and control work, there are problems such as insufficient accuracy of investigation and evaluation, inadequate application of advanced means, and weak ability of hidden danger identification and monitoring and early warning. Taking Zigui County in the hinterland of Three Gorges Reservoir Area, China as the study area, we develop quantitative and qualitative analysis methods of multi-source data and multi-model fusion. Focusing on the early identification of landslide hazards under the constraints of unstable slope unit and effective deformation aggregation area, we realize the early identification of landslide hazards based on the SBAS-InSAR technology and deep neural network model.

Methods First, we construct a set of landslide early identification indexes covering landslide susceptibility, slope unit and surface deformation information, and comprehensively evaluate the applicability of three data balancing methods, namely, synthetic minority over-sampling technique (SMOTE), Tomek Link under-sampling and SMOTETomek integrated sampling, as well as three deep neural network models, namely, convolutional neural network, multilayer perceptron (MLP) and long short-term memory network, to evaluate landslide susceptibility and extract unstable slope units. Second, we use small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) to invert the surface deformation rate in the study area, and obtain the real deformation results in the slope direction based on the terrain visibility analysis and the maximum slope projection. Finally, we couple the unstable slope units with the surface deformation results to delineate the potential landslide areas, superimpose them on Google Earth for visual interpretation for the early identification of landslides, and verify them by comparing them with the field investigation data.

Results (1) The joint MLP-Tomek Link model outperforms other models in terms of accuracy, precision, and F1 score, and is the best landslide susceptibility assessment model in the region. (2) The effective deformation results for 2019—2022 in Zigui County are concentrated in the north-central and southeastern tectonic fault zones, and the annual average deformation rate in the direction of its maximum slope ranges from 0-112.68 mm/a. (3) A total of 30 potential landslide zones are screened, and landslide hazards are successfully identified in 21 of them, with an accuracy of 70% in the early landslide identification.

Conclusions The combination of SBAS-InSAR technology and deep neural network model can significantly improve the accuracy of early landslide identification, providing technical and methodological references for landslide risk evaluation and management, landslide monitoring, early warning and prevention.