Numerical Recipes for Landslide Spatial Prediction Using R-INLA: A Step-by-Step Tutorial

L. Lombardo, Raphaël Huser, Thomas Opitz

Research output: Chapter in Book/Report/Conference proceedingChapterAcademicpeer-review

Abstract

The geomorphological community typically assesses the landslide susceptibility at the catchment or larger scales through spatial predictive models. However, the spatial information is conveyed only through the geographical distribution of the covariates. Spatial dependence, which allows capturing similarities at neighboring sites that are not directly explained by covariate information, is typically not accounted for in the landslides literature, whilst such spatial models have become commonplace in the geostatistical literature. Here we explain step by step how to rigorously model and predict activations of debris flow based on an adequate statistical model by using the R-INLA library from the statistical software R in the context of a real multiple landslide event. This chapter follows the analysis of Lombardo, Opitz, and Huser with a few modifications; it is written in a tutorial style to provide the geomorphological community with a hands-on guide to replicate similar analyses in R. While our focus here is on implementation and computing, more details about the underlying statistical theory, modeling, and estimation can be found in the original work by Lombardo, Opitz, and Huser Our modeling approach deviates fundamentally from the commonly used regression models fitted to binary presence/absence data. Specifically, we use a Bayesian hierarchical Cox point process model to describe landslide counts at high resolution (i.e., at the pixel level), while capturing spatial dependence through a latent spatial effect defined at lower resolution over slope units. Our point process modeling approach allows us to derive the distribution of aggregated landslide counts for any areas of interest. Crucially, the latent spatial effect represents the unexplained but spatially structured component in the data when the linear or nonlinear effects of covariates are removed. Thus, in the case of sparse rain gauge or seismic networks, we suggest using the latent spatial effect to uncover the trigger distribution over space. In particular, for landslides triggered by extreme precipitation, the meteorological stress can play a dominant role with respect to the covariates that are typically introduced in predictive models; therefore, accounting for the trigger in modeling may dramatically improve the performance of landslide prediction.
Original languageEnglish
Title of host publicationSpatial Modeling in GIS and R for Earth and Environmental Sciences
EditorsH.R. Pourghasemi, C. Gokceoglu
PublisherElsevier
Chapter3
Pages55-83
Number of pages29
ISBN (Electronic)978-0-12-815226-3
ISBN (Print)978-0-12-815226-3
DOIs
Publication statusPublished - 2019

Fingerprint

landslide
prediction
modeling
debris flow
geographical distribution
gauge
pixel
catchment
software
effect

Keywords

  • Cox Point Process
  • Integrated Nested Laplace Approximation (INLA)
  • Landslide Susceptibility
  • Landslide Intensity
  • R-INLA
  • Slope Unit
  • Spatial Point Pattern

Cite this

Lombardo, L., Huser, R., & Opitz, T. (2019). Numerical Recipes for Landslide Spatial Prediction Using R-INLA: A Step-by-Step Tutorial. In H. R. Pourghasemi, & C. Gokceoglu (Eds.), Spatial Modeling in GIS and R for Earth and Environmental Sciences (pp. 55-83). Elsevier. https://doi.org/10.1016/B978-0-12-815226-3.00003-X
Lombardo, L. ; Huser, Raphaël ; Opitz, Thomas. / Numerical Recipes for Landslide Spatial Prediction Using R-INLA : A Step-by-Step Tutorial. Spatial Modeling in GIS and R for Earth and Environmental Sciences. editor / H.R. Pourghasemi ; C. Gokceoglu. Elsevier, 2019. pp. 55-83
@inbook{b0392d06f2554b56a03aafad230ea60f,
title = "Numerical Recipes for Landslide Spatial Prediction Using R-INLA: A Step-by-Step Tutorial",
abstract = "The geomorphological community typically assesses the landslide susceptibility at the catchment or larger scales through spatial predictive models. However, the spatial information is conveyed only through the geographical distribution of the covariates. Spatial dependence, which allows capturing similarities at neighboring sites that are not directly explained by covariate information, is typically not accounted for in the landslides literature, whilst such spatial models have become commonplace in the geostatistical literature. Here we explain step by step how to rigorously model and predict activations of debris flow based on an adequate statistical model by using the R-INLA library from the statistical software R in the context of a real multiple landslide event. This chapter follows the analysis of Lombardo, Opitz, and Huser with a few modifications; it is written in a tutorial style to provide the geomorphological community with a hands-on guide to replicate similar analyses in R. While our focus here is on implementation and computing, more details about the underlying statistical theory, modeling, and estimation can be found in the original work by Lombardo, Opitz, and Huser Our modeling approach deviates fundamentally from the commonly used regression models fitted to binary presence/absence data. Specifically, we use a Bayesian hierarchical Cox point process model to describe landslide counts at high resolution (i.e., at the pixel level), while capturing spatial dependence through a latent spatial effect defined at lower resolution over slope units. Our point process modeling approach allows us to derive the distribution of aggregated landslide counts for any areas of interest. Crucially, the latent spatial effect represents the unexplained but spatially structured component in the data when the linear or nonlinear effects of covariates are removed. Thus, in the case of sparse rain gauge or seismic networks, we suggest using the latent spatial effect to uncover the trigger distribution over space. In particular, for landslides triggered by extreme precipitation, the meteorological stress can play a dominant role with respect to the covariates that are typically introduced in predictive models; therefore, accounting for the trigger in modeling may dramatically improve the performance of landslide prediction.",
keywords = "Cox Point Process, Integrated Nested Laplace Approximation (INLA), Landslide Susceptibility, Landslide Intensity, R-INLA, Slope Unit, Spatial Point Pattern",
author = "L. Lombardo and Rapha{\"e}l Huser and Thomas Opitz",
year = "2019",
doi = "10.1016/B978-0-12-815226-3.00003-X",
language = "English",
isbn = "978-0-12-815226-3",
pages = "55--83",
editor = "H.R. Pourghasemi and C. Gokceoglu",
booktitle = "Spatial Modeling in GIS and R for Earth and Environmental Sciences",
publisher = "Elsevier",

}

Lombardo, L, Huser, R & Opitz, T 2019, Numerical Recipes for Landslide Spatial Prediction Using R-INLA: A Step-by-Step Tutorial. in HR Pourghasemi & C Gokceoglu (eds), Spatial Modeling in GIS and R for Earth and Environmental Sciences. Elsevier, pp. 55-83. https://doi.org/10.1016/B978-0-12-815226-3.00003-X

Numerical Recipes for Landslide Spatial Prediction Using R-INLA : A Step-by-Step Tutorial. / Lombardo, L.; Huser, Raphaël; Opitz, Thomas.

Spatial Modeling in GIS and R for Earth and Environmental Sciences. ed. / H.R. Pourghasemi; C. Gokceoglu. Elsevier, 2019. p. 55-83.

Research output: Chapter in Book/Report/Conference proceedingChapterAcademicpeer-review

TY - CHAP

T1 - Numerical Recipes for Landslide Spatial Prediction Using R-INLA

T2 - A Step-by-Step Tutorial

AU - Lombardo, L.

AU - Huser, Raphaël

AU - Opitz, Thomas

PY - 2019

Y1 - 2019

N2 - The geomorphological community typically assesses the landslide susceptibility at the catchment or larger scales through spatial predictive models. However, the spatial information is conveyed only through the geographical distribution of the covariates. Spatial dependence, which allows capturing similarities at neighboring sites that are not directly explained by covariate information, is typically not accounted for in the landslides literature, whilst such spatial models have become commonplace in the geostatistical literature. Here we explain step by step how to rigorously model and predict activations of debris flow based on an adequate statistical model by using the R-INLA library from the statistical software R in the context of a real multiple landslide event. This chapter follows the analysis of Lombardo, Opitz, and Huser with a few modifications; it is written in a tutorial style to provide the geomorphological community with a hands-on guide to replicate similar analyses in R. While our focus here is on implementation and computing, more details about the underlying statistical theory, modeling, and estimation can be found in the original work by Lombardo, Opitz, and Huser Our modeling approach deviates fundamentally from the commonly used regression models fitted to binary presence/absence data. Specifically, we use a Bayesian hierarchical Cox point process model to describe landslide counts at high resolution (i.e., at the pixel level), while capturing spatial dependence through a latent spatial effect defined at lower resolution over slope units. Our point process modeling approach allows us to derive the distribution of aggregated landslide counts for any areas of interest. Crucially, the latent spatial effect represents the unexplained but spatially structured component in the data when the linear or nonlinear effects of covariates are removed. Thus, in the case of sparse rain gauge or seismic networks, we suggest using the latent spatial effect to uncover the trigger distribution over space. In particular, for landslides triggered by extreme precipitation, the meteorological stress can play a dominant role with respect to the covariates that are typically introduced in predictive models; therefore, accounting for the trigger in modeling may dramatically improve the performance of landslide prediction.

AB - The geomorphological community typically assesses the landslide susceptibility at the catchment or larger scales through spatial predictive models. However, the spatial information is conveyed only through the geographical distribution of the covariates. Spatial dependence, which allows capturing similarities at neighboring sites that are not directly explained by covariate information, is typically not accounted for in the landslides literature, whilst such spatial models have become commonplace in the geostatistical literature. Here we explain step by step how to rigorously model and predict activations of debris flow based on an adequate statistical model by using the R-INLA library from the statistical software R in the context of a real multiple landslide event. This chapter follows the analysis of Lombardo, Opitz, and Huser with a few modifications; it is written in a tutorial style to provide the geomorphological community with a hands-on guide to replicate similar analyses in R. While our focus here is on implementation and computing, more details about the underlying statistical theory, modeling, and estimation can be found in the original work by Lombardo, Opitz, and Huser Our modeling approach deviates fundamentally from the commonly used regression models fitted to binary presence/absence data. Specifically, we use a Bayesian hierarchical Cox point process model to describe landslide counts at high resolution (i.e., at the pixel level), while capturing spatial dependence through a latent spatial effect defined at lower resolution over slope units. Our point process modeling approach allows us to derive the distribution of aggregated landslide counts for any areas of interest. Crucially, the latent spatial effect represents the unexplained but spatially structured component in the data when the linear or nonlinear effects of covariates are removed. Thus, in the case of sparse rain gauge or seismic networks, we suggest using the latent spatial effect to uncover the trigger distribution over space. In particular, for landslides triggered by extreme precipitation, the meteorological stress can play a dominant role with respect to the covariates that are typically introduced in predictive models; therefore, accounting for the trigger in modeling may dramatically improve the performance of landslide prediction.

KW - Cox Point Process

KW - Integrated Nested Laplace Approximation (INLA)

KW - Landslide Susceptibility

KW - Landslide Intensity

KW - R-INLA

KW - Slope Unit

KW - Spatial Point Pattern

UR - https://ezproxy2.utwente.nl/login?url=https://library.itc.utwente.nl/login/2019/chap/lombardo_num.pdf

U2 - 10.1016/B978-0-12-815226-3.00003-X

DO - 10.1016/B978-0-12-815226-3.00003-X

M3 - Chapter

SN - 978-0-12-815226-3

SP - 55

EP - 83

BT - Spatial Modeling in GIS and R for Earth and Environmental Sciences

A2 - Pourghasemi, H.R.

A2 - Gokceoglu, C.

PB - Elsevier

ER -

Lombardo L, Huser R, Opitz T. Numerical Recipes for Landslide Spatial Prediction Using R-INLA: A Step-by-Step Tutorial. In Pourghasemi HR, Gokceoglu C, editors, Spatial Modeling in GIS and R for Earth and Environmental Sciences. Elsevier. 2019. p. 55-83 https://doi.org/10.1016/B978-0-12-815226-3.00003-X