Landslide Negative Sample Construction and Susceptibility Assessment Based on the Temporal Stability of Dual SAR Indices: A Case Study of the Lijiang River Basin

Selecting reliable negative samples (NS) is a crucial step in enhancing the robustness of landslide susceptibility assessments (LSA). Existing studies frequently utilize InSAR-derived deformation data to identify stable areas, thereby refining the negative samples; however, InSAR is susceptible to residual errors due to decorrelation and geometric distortions, particularly in regions with significant topographic relief and dense vegetation cover. Additionally, the processing workflow for InSAR can be complex and costly.

To address these issues, we examine the Lijiang River Basin in Guangxi, China, as a case study. We propose a novel negative sampling strategy constrained by the temporal stability of two SAR-based indices: Long-term Distillation and Identification (LDI). First, we delineate temporally stable areas (S_SAR) by selecting pixels that exhibit minimal long-term change rates in the Radar Vegetation Index (RVI) and the Radar Forest Degradation Index (RFDI). We then apply Positive-Unlabeled Learning (PU-learning) to refine S_SAR further, resulting in a high-confidence NS set (Nopt). Next, we evaluate stability differences between Nopt and various NS sets generated by conventional sampling strategies, using cumulative deformation and deformation-rate metrics obtained from SBAS-InSAR. Finally, we built Landslide Susceptibility Assessment (LSA) models utilizing Random Forest (RF), Extreme Gradient Boosting (Xgboost), and Categorical Boosting (Catboost). We assess model performance using the Area Under the Curve (AUC) and confusion-matrix-based metrics. Additionally, we analyze spatial patterns in LSA, area proportions across susceptibility classes, and their relationship with the multi-year means and long-term change rates of RVI and RFDI.

The results indicate the following: (1) Deformation values in S_SAR are primarily clustered around 0 mm, confirming the consistency between “stable long-term vegetation change” and “stable ground deformation.” After refining with PU-learning, Nopt shows more minor fluctuations in deformation and exhibits the highest internal consistency. (2) LSA models based on LDI perform the best, with the Xgboost-based LSA achieving the highest AUC (0.843). Additionally, feature contributions quantified by Shapley Additive Explanations (SHAP) are more concentrated and stable, demonstrating that LDI effectively reduces noise. (3) Although various NS sampling strategies result in significant differences in LSA spatial patterns, the Very High Susceptibility (VH) class consistently displays a typical pattern of “higher RFDI and lower RVI, with a weaker RFDI trend and a stronger RVI trend”. This suggests that areas classified as VH have lower vegetation cover, greater inter-annual variability, and weaker disturbance resistance. Overall, LDI provides a cost-effective approach to obtaining reliable NS data in complex terrains, serving as a valuable reference for LSA modeling in the Lijiang River Basin and similar regions.