A new method of landslide susceptibility assessment considering time-series InSAR surface deformation optimized non-landslide sample

Due to the severe casualties and property losses caused by landslides, carrying out landslide susceptibility assessment (LSA) has important theoretical value and practical significance. However, affected by factors such as non-landslide sample selection and prediction model parameter setting, the results of landslide susceptibility evaluation have a certain deviation from the actual situation. Based on the dual sampling strategy of non-landslide ‘low deformation’ and ‘buffer zone’, this paper uses the Northern Goshawk Optimization (NGO) algorithm to enhance the Random Forest (RF) model, and proposes an NGO-RF susceptibility evaluation method considering the optimization of non-landslide samples with time-series InSAR deformation. Taking Shuicheng District, Guizhou Province, China as the study area, the evaluation results are compared and analyzed with the RF, CatBoost and Multiple Linear Regression (MLR) models. The research results show that the evaluation results under the ‘low deformation’ sampling strategy are better than those under the ‘buffer zone’ strategy. In addition, the AUC-ROC value of the NGO-RF model under the ‘low deformation’ sampling strategy is 0.939, which is at least 2.3% higher than that of the RF, CatBoost and MLR models, presenting better evaluation performance. Meanwhile, different models also show certain differences in predicting the distribution of historical landslide points in the ‘low deformation’ and ‘buffer zone’ scenarios. The number of landslide points in the ‘very high’ risk zone predicted by the NGO-RF model increases from 86 to 136, while the number of landslide points in the ‘low’ risk zone decreases by 5, which significantly improves the accuracy and reliability of high-risk zones. This effectively verifies the effectiveness of the time-series InSAR deformation‑optimized non-landslide sample selection strategy, and can provide technical support for landslide disaster prediction, early warning, disaster prevention and mitigation.