Exploring the uncertainty of landslide susceptibility assessment caused by the number of non–landslides

Identifying the uncertainty caused by the number of non-landslides is necessary to obtain a precise landslide susceptibility map. Hence, the objective of this study is to investigate the evolution of landslide susceptibility under different numbers of non-landslides, and thus to answer the question of how many non-landslides are required for susceptibility modeling. For this purpose, Heilongjiang province was selected as the study area to build the landslide inventory, and the number of non-landslides at different ratios to landslides was set. Later, 12 conditioning factors were constructed and applied to the susceptibility model. Finally, the susceptibility uncertainty was ascertained via fluctuations in sample prediction, variations in classifier performance, and changes in the susceptibility map. Results show that the fluctuation in landslide prediction grows, while that for non-landslides decreases, as the number of non-landslides increases. Quantitatively, the sensitivity is 0.93 at 0.5/1, followed by 0.90 at 1/1, 0.87 at 2/1, 0.72 at 5/1, and 0.61 at 10/1; the specificity is 0.86 at 0.5/1, followed by 0.90 at 1/1, 0.92 at 2/1, 0.97 at 5/1, and 0.99 at 10/1; while the accuracy is 0.91 at 0.5/1, followed by 0.90 at 1/1, 0.89 at 2/1, 0.84 at 5/1, and 0.96 at 10/1, in the training phase. Also, the predictive ability of samples shows similar trends, in the test phase. Besides, the area with high susceptibility gradually shrinks with increasing non-landslides, and the R-value reaches the top, 2.92, at the equilibrium state. This study provides an in-depth understanding to emphasize the balanced value for non-landslides when applying the idea of presence-absence.