Integrating empirical models and satellite radar can improve landslide detection for emergency response

Katy Burrows,Richard J Walters,Dino Bellugi,David Milledge

doi:10.5194/nhess-21-2993-2021

Abstract

Abstract. Information on the spatial distribution of triggered landslides following an earthquake is invaluable to emergency responders. Manual mapping using optical satellite imagery, which is currently the most common method of generating this landslide information, is extremely time consuming and can be disrupted by cloud cover. Empirical models of landslide probability and landslide detection with satellite radar data are two alternative methods of generating information on triggered landslides that overcome these limitations. Here we assess the potential of a combined approach, in which we generate an empirical model of the landslides using data available immediately following the earthquake using the random forest technique and then progressively add landslide indicators derived from Sentinel-1 and ALOS-2 satellite radar data to this model in the order they were acquired following the earthquake. We use three large case study earthquakes and test two model types: first, a model that is trained on a small part of the study area and used to predict the remainder of the landslides and, second, a preliminary global model that is trained on the landslide data from two earthquakes and used to predict the third. We assess model performance using receiver operating characteristic analysis and r2, and we find that the addition of the radar data can considerably improve model performance and robustness within 2 weeks of the earthquake. In particular, we observed a large improvement in model performance when the first ALOS-2 image was added and recommend that these data or similar data from other L-band radar satellites be routinely incorporated in future empirical models.

Highlights

Earthquake-triggered landslides are a major secondary hazard associated with large continental earthquakes and disrupt emergency response efforts
We have demonstrated that the addition of InSAR coherence features (ICFs) to empirical models of landslide areal density (LAD) based on topography, ground shaking estimates, land cover and lithology significantly improves their performance in terms of receiver operating characteristic (ROC) area under this curve (AUC) and r2
The performance of all models was better with ROC analysis than with r2, indicating that the models are better suited to discriminating between landslide and nonlandslide areas following an earthquake than predicting continuous landslide areal density

Summary

Introduction

Earthquake-triggered landslides are a major secondary hazard associated with large continental earthquakes and disrupt emergency response efforts. Information on their spatial distribution is required to inform this emergency response but must be generated within 2 weeks of the earthquake in order to be most useful (Inter-Agency Standing Committee, 2015; Williams et al, 2018). The most common method of generating landslide information is manual mapping using optical satellite imagery, but this is a time-consuming process and can be delayed by weeks or even months due to cloud cover (Robinson et al, 2019), leading to incomplete landslide information during the emergency response. The second is to estimate landslide locations based on their signal in satellite synthetic aperture radar (SAR) data, which can be acquired through cloud cover and so is often able to provide more complete spatial coverage than optical satellite imagery in the critical 2-week response window (e.g. Aimaiti et al, 2019; Burrows et al, 2019, 2020; Jung and Yun, 2019; Konishi and Suga, 2019; Mondini et al, 2019, 2021)

Objectives

Results

Discussion

Conclusion