Abstract: Due to the fact that the downstream of the Wudongde hydropower station is located in the Jinsha River basin, the flood discharge of the hydropower station will cause changes in the water level in the downstream area and frequent geological activities near the fault zone, thereby changing the stability of the slopes around the basin. It is highly prone to landslide disasters and causing river blockage. In order to gain the understanding of the slope conditions in high and steep mountainous areas, this experiment utilizes 81 Sentinel-A images from January 2021 to October 2023, and obtains surface deformation characteristics using SBAS-InSAR technology. By combining deformation rate maps and remote sensing images, potential landslide risk areas are identified. A1, A2, and A4 are selected as typical areas and analyzed through cumulative temporal deformation. A LSTM neural network settlement prediction model is constructed, and then the GWO algorithm for hyperparameter optimization is used. The settlement data obtained from the final point selection is divided into a training set and a testing set, and compared with the accuracy of the traditional prediction model SVR. The results show that the GWO-LSTM model has high accuracy in predicting landslide deformation in complex mountainous areas. Among the 9 test points, the maximum mean absolute error is 1.080 8 mm and the maximum root mean square error is 1.194 2 mm. This paper provides a theoretical basis for landslide disaster warning and management.

Key words: SABS-InSAR, landslide identification, GWO, LSTM memory network