Loess distributed across the Loess Plateau is characterized by prominent water sensitivity, collapsibility, and well-developed vertical joints, making regional landslides frequently triggered by rainfall infiltration, freeze-thaw cycles, and intensive human engineering activities. With the large-scale construction of ultra-high-voltage power transmission infrastructure in mountainous loess terrain, refined landslide susceptibility assessment has become an essential prerequisite for engineering safety management. However, remote hilly loess regions generally suffer from incomplete historical landslide inventories. Conventional sampling strategies obtain positive samples merely from archived landslide records and extract negative samples randomly across the entire study area. Such sampling patterns lead to insufficient positive samples and contaminated negative samples mixed with ambiguous non-landslide grids that share similar geological settings, which seriously degrades the prediction performance of machine learning-based susceptibility models. To solve this technical bottleneck, this study proposes a collaborative optimization strategy for positive and negative landslide samples by integrating time-series small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) remote sensing interpretation and the information value method.

The study area was located at the southern foot of Lyuliang Mountain in Linfen, Shanxi Province, covering a total area of 189.62 km² with typical loess ridge-gully geomorphology. Nine assessment factors closely related to loess landslide initiation were selected for susceptibility modeling: elevation, slope gradient, slope aspect, plan curvature, profile curvature, topographic wetness index (TWI), normalized difference vegetation index (NDVI), gully density, and distance to gullies. Based on Sentinel-1A ascending SAR images collected from March 2023 to June 2024, SBAS-InSAR deformation inversion was implemented, and grid cells with annual sliding deformation rate ≤ −15 mm/a were preliminarily defined as potential unstable landslide zones. Combined with visual interpretation of typical geomorphic features such as cirque-shaped scarps from high-resolution optical remote sensing images, dual verification was conducted to screen reliable positive samples. Specifically, 230 raster positive samples were expanded from 10 historically recorded landslides, and another 920 supplementary raster samples were identified from 31 newly detected hidden landslides, forming a final positive dataset consisting of 1 150 grid cells. Subsequently, the information value model was adopted to classify the entire study area into five susceptibility grades via the natural breaks algorithm, and qualified negative samples were randomly selected only from extremely low and low susceptibility zones with a fixed 1:1 positive-to-negative sample ratio. Four comparative sampling schemes were constructed for quantitative comparison, and all datasets were randomly split into training and testing subsets at a ratio of 7:3. Random forest (RF) and back propagation neural network (BP) were employed to establish landslide susceptibility models, and the area under the receiver operating characteristic curve (ROC-AUC) was adopted as the quantitative assessment indicator of model accuracy.

The modeling results revealed an obvious hierarchical improvement effect. Optimizing only positive samples greatly improved model accuracy, with AUC values reaching 0.87608 (RF) and 0.77174 (BP), while independent negative sample optimization brought limited accuracy improvement, with AUC values of 0.59124 (RF) and 0.58785 (BP). The collaborative optimization scheme combining remote-sensing-derived positive samples and information-value-filtered negative samples achieved optimal performance, with RF-AUC = 0.91812 and BP-AUC = 0.81937, representing accuracy improvements of 60.27% and 47.43%, respectively, compared with the traditional sampling scheme (RF-AUC = 0.57285, BP-AUC = 0.55577).

This study verifies that the proposed hybrid sample optimization framework can significantly improve the reliability of loess landslide susceptibility assessment. The core technical idea can be extended to other data-deficient regions such as red-bed hilly terrains and alpine canyon areas, providing solid technical support for geological disaster prevention and safe operation of major power transmission projects on the Loess Plateau.