A New Negative-Sample Construction Strategy for Landslide Susceptibility Mapping

This is the authors' abstract. We don't add key points for this paper.

Negative-sample construction is an important step in machine learning-based landslide susceptibility assessment. Traditional negatives selected from flat or clearly stable terrain may introduce easy-negative bias and inflate apparent model performance. In this study, positive-unlabeled (PU) bagging and hard-negative mining were used to construct a mixed negative set combining easy negatives with harder negatives from plausible hillslope environments. To evaluate the effects of negative-sample construction, a dual-branch deep learning model comprising a hydrological sequence branch and a static environmental branch was used to assess landslide susceptibility in Kentucky, USA. The model integrated antecedent precipitation and terrestrial water-storage anomaly sequences with topographic, soil, vegetation, and lithological predictors. Five scenarios represented different negative-sample characteristics and levels of environmental matching: easy-negative, regular-mixed, hard-negative, slope-matched, and static-matched. Each scenario was applied to models trained with traditional and reconstructed negatives to compare their performance under different negative-sample conditions. The reconstructed strategy achieved higher AUC values in the hard-negative and slope-matched scenarios (0.823 and 0.752, respectively) than the traditional strategy (0.786 and 0.659, respectively) and retained higher recall at a false-positive rate (FPR) of 0.10. Conventional machine learning models showed the same pattern, indicating that the improvement was not architecture-specific. These findings support more representative negative-sample construction and more reliable regional model evaluation, providing a stronger basis for risk-informed land-use planning and landslide mitigation.