Effective surveyed area and its role in statistical landslide susceptibility assessments

Txomin Bornaetxea,Mauro Rossi,Ivan Marchesini,Massimiliano Alvioli

doi:10.5194/nhess-18-2455-2018

Abstract

Abstract. Geomorphological field mapping is a conventional method used to prepare landslide inventories. The approach is typically hampered by the accessibility and visibility, during field campaigns for landslide mapping, of the different portions of the study area. Statistical significance of landslide susceptibility maps can be significantly reduced if the classification algorithm is trained in unsurveyed regions of the study area, for which landslide absence is typically assumed, while ignorance about landslide presence should actually be acknowledged. We compare different landslide susceptibility zonations obtained by training the classification model either in the entire study area or in the only portion of the area that was actually surveyed, which we name effective surveyed area. The latter was delineated by an automatic procedure specifically devised for the purpose, which uses information gathered during surveys, along with landslide locations. The method was tested in Gipuzkoa Province (Basque Country), north of the Iberian Peninsula, where digital thematic maps were available and a landslide survey was performed. We prepared the landslide susceptibility maps and the associated uncertainty within a logistic regression model, using both slope units and regular grid cells as the reference mapping unit. Results indicate that the use of effective surveyed area for landslide susceptibility zonation is a valid approach that minimises the limitations stemming from unsurveyed regions at landslide mapping time. Use of slope units as mapping units, instead of grid cells, mitigates the uncertainties introduced by training the automatic classifier within the entire study area. Our method pertains to data preparation and, as such, the relevance of our conclusions is not limited to the logistic regression but are valid for virtually all the existing multivariate landslide susceptibility models.

Highlights

Landslide susceptibility is defined as the likelihood of a landslide occurring in an area on the basis of the local terrain and environmental conditions (Brabb, 1984; Guzzetti et al, 2005)
We defined the obtained results as whole-area pixel map (WAPM) and effective surveyed-area pixel map (ESA-PM). In both whole area (WA)-PM and effective surveyed area (ESA)-PM, we first used the same 13 explanatory variables listed in Table 2, and we selected for each model assessment the most relevant explanatory variables considering the collinearity between each pair of variables and the significance (p value) of the regression estimates
Inspection of the four-fold or contingency plots (Fig. 3a, d) reveals that WA-PM correctly predicted the 63.58 % (TP+true negatives (TN)) of the observed unstable and stable mapping units, whereas ESA-PM was capable of correctly predicting a higher number of mapping units (65.45 %)

Summary

Introduction

Landslide susceptibility is defined as the likelihood of a landslide occurring in an area on the basis of the local terrain and environmental conditions (Brabb, 1984; Guzzetti et al, 2005). Landslide susceptibility zonation (LSZ) is important for landslide mitigation plans, since it supplies planners and decision makers with essential information (Van Den Eeckhaut et al, 2012). Many statistical methods, aimed at estimating the propensity of a territory to experience slope failures, rely on landslide inventory maps and spatial thematic layers as predisposing factors (Ermini et al, 2005; Van Den Eeckhaut et al, 2006; Camilo et al, 2017). In statistical landslide susceptibility models, such as the logistic regression (LR) model adopted in this work, the preparation of the training data set is a fundamental and critical step. This requires the selection of a sample of stable (without landslides) and unstable (with landslides) mapping units.

Objectives

Results

Discussion

Conclusion