Influence of management of variables, sampling zones and land units on LR analysis for landslide spatial prevision
Several authors, according to different methodological approaches, have employed logistic Regression (LR), a multivariate statistical analysis adopted to assess the spatial probability of landslide, even though its fundamental principles have remained unaltered.
This study aims at assessing the influence of some of these methodological approaches on the performance of LR, through a series of sensitivity analyses developed over a test area of about 300 km 2 in Calabria (southern Italy).
In particular, four types of sampling (1 – the whole study area; 2 – transects running parallel to the general slope direction of the study area with a total surface of about 1/3 of the whole study area; 3 – buffers surrounding the phenomena with a 1/1 ratio between the stable and the unstable area; 4 – buffers surrounding the phenomena with a 1/2 ratio between the stable and the unstable area), two variable coding modes (1 – grouped variables; 2 – binary variables), and two types of elementary land (1 – cells units; 2 – slope units) units have been tested. The obtained results must be considered as statistically relevant in all cases (Aroc values > 70%), thus confirming the soundness of the LR analysis which maintains high predictive capacities notwithstanding the features of input data.
As for the area under investigation, the best performing methodological choices are the following: (i) transects produced the best results (0 < P(y) ≤ 93.4%; Aroc = 79.5%); (ii) as for sampling modalities, binary variables (0 < P(y) ≤ 98.3%; Aroc = 80.7%) provide better performance than ordinated variables; (iii) as for the choice of elementary land units, slope units (0 < P(y) ≤ 100%; Aroc = 84.2%) have obtained better results than cells matrix.