An improved information quantity method for non-landslide selection to enhance landslide susceptibility evaluation: a case study in Yongfeng, South China

Machine learning models are extensively employed in landslide susceptibility evaluation, where the selection of non-landslide samples significantly impacts model performance. Traditional methods for non-landslide sample selection often yield inaccurate landslide susceptibility maps due to limitations in spatial representativeness and sample reliability. In this study, an improved information quantity method (IIQM) was proposed to evaluate landslide susceptibility in Yongfeng, South China. Specifically, non-landslide samples were randomly generated in areas with very low and low susceptibility, at least 1,200 m away from any other landslide sample. Eight influential factors were selected and analyzed, including slope, elevation, drainage networks, NDVI, roads, land-use types, lithology, and faults. Sample sets of historical landslides and non-landslides derived from the random selection method (RSM), information quantity method (IQM), and IIQM were divided into training and validation sets with a 4:1 ratio. Landslide susceptibility maps were generated using information quantity-logistic regression and respective training sets. The performance of the landslide susceptibility models was evaluated via specificity, sensitivity, accuracy, and area under the curve (AUC). Results demonstrated that IIQM-based models outperformed RSM and IQM, achieving a median AUC of 0.87 and an average AUC of 0.88, compared to RSM (median AUC of 0.68 and average AUC of 0.64) and IQM (median AUC of 0.86 and average AUC of 0.85). The IIQM effectively mitigates the inclusion of false non-landslides and improves model generalization, offering a robust framework for non-landslide selection in machine learning-based landslide susceptibility assessments. Also, proximity to roads (0–200 m, information value = 1.04) and lithology (hard-soft integrated stones, InV = 0.54) were identified as dominant contributors to landslide susceptibility in Yongfeng. These findings provide valuable insights into the selection of non-landslide samples, particularly within the context of information quantity-logistic regression and other machine learning models, and may also provide scientific reference for the regional landslide prevention.