Motivation
The Asahi Shimbun/Getty Images
• Landslides are a major natural hazard and are often the dominant process that
erode mountainous landscapes
• Rapid response to landslide events (and other natural hazards) is necessary to
assess damages and save lives
• Efforts are impeded by a lack of detailed information on the condition or
location of damaged areas
• Construct inventories to improve understanding of where landslides occur, to
quantify erosion, and to look for areas subject to secondary hazards
• It is necessary to continue to develop easy-to-use tools with freely available
data that can be used to map the landslide extent and level of damage
following catastrophic events
Outline
Image from PlanetLabs
• Introduction
• Landslide mapping
• Data and Methods
• Synthetic Aperture Radar
• Landslide identification in Google Earth Engine
• Results (Case study: 2018 Hiroshima, Japan)
• Landslide Detection with SAR
• Rapid Response with SAR
• Summary
• Examples
• 2018 earthquake-triggered landslides Hokkaido, Japan
• 2018 post-fire debris flows, Montecito, CA, USA
• 2020 Hurricane Eta landslides, Guatemala and Honduras
Landslide Identification
1 km
Image from Sentinel-2
Satellite-based Optical Imagery
2018 earthquake-triggered landslides Hokkaido, Japan
• Pixel sizes from centimeters to hundreds of meters
• Collected every few days to weeks
• Multiple spectral bands (visible and near-infrared)
Landslide Identification
Satellite-based Optical Imagery
1 km
2018 Hokkaido, Japan
Image from Sentinel-2
• Require solar illumination
• Issues caused by clouds,
shadows, night time
Limitations
• Pixel sizes from centimeters to hundreds of meters
• Collected every few days to weeks
• Multiple spectral bands (visible and near-infrared)
Image from Sentinel-2
2018 earthquake-triggered landslides Hokkaido, Japan
Landslide Identification
• Normalized difference vegetation index (NDVI)
• NDVI change detection methods have been
used for identifying landslides
(e.g., Huang et al., 2020; Mondini et al., 2011; Scheip and Wegmann, 2020)
Satellite-based Optical Imagery
Scheip and Wegmann, 2020
1 km
2019 rainfall-induced landslides,
West Pokot County, Kenya
Vegetation
change
(%rdNDVI)
gain
loss
• Require solar illumination
• Issues caused by clouds,
shadows, night time
Limitations
Landslide Identification
• 1D viewing geometry
• Topography
• Vegetation
Limitations
• Pixel sizes meters to tens of meters
• Collected every few days to weeks
• Operates day or night and can ”see” through clouds
• Amplitude and Phase data
Satellite-based Synthetic Aperture Radar (SAR)
Images from Sentinel 1 and 2
2018 earthquake-triggered landslides Hokkaido, Japan
Landslide Identification
• Damage mapping from interferometric coherence (phase data)
• Measure the change in radar scattering characteristics
2018 earthquake-triggered landslides Hokkaido, Japan
Jung and Yun, 2020
Satellite-based Synthetic Aperture Radar (SAR)
Landslide Identification
• Damage mapping from amplitude change
• Measure the change backscatter brightness
Satellite-based Synthetic Aperture Radar (SAR)
Jung and Yun, 2020
2018 earthquake-triggered landslides Hokkaido, Japan
We will use a variation of this approach today
• Backscattering coefficient is strength
of reflected echo, often expressed in
decibels (dB)
• Backscatter depends on properties of
ground
• E.g., soil moisture, surface roughness,
vegeation
• Can use SAR backscatter to detect a
variety of natural disasters such as
floods, forest fires, earthquakes,
landslides
(Burrows et al., in review; DeVries et al., 2020; Jung
and Yun, 2020; Mondini et al., 2019; Rignot and Van
Zyl, 1993; Tay et al., 2020; Yun et al., 2015)
Source: ESA Earth Online
Methods
Slide from S. Jones
SAR Backscatter
• Smooth surfaces: scatters energy away and appears dark (e.g., open water)
• Rough surfaces: scatters energy away and towards satellite
• Volume scattering: scatters multiple times
• Double Bounce: scatters back toward satellite, common in urban area
Methods
Slide from E. Podest SAR for Flood Mapping Using Google Earth Engine
SAR Backscatter Image from Ulaby et al., 1981
• Longer radar wavelengths can better penetrate through vegetation
and even into the shallow ground surface
Methods
Slide from E. Podest SAR for Flood Mapping Using Google Earth Engine
SAR Backscatter
Methods
Slide from E. Podest SAR for Flood Mapping Using Google Earth Engine
SAR Backscatter
Speckle
• Each cell contains several scattering centers whose
returns create positive or negative interference
• "salt and pepper" appearance
• Corrected with smoothing filters and multi-looking
Methods
SAR Backscatter Issues
Source:
ESA
Geometric Issues
• Shadowing
• Foreshortening (slope facing toward
radar)
• Layover (mountains look as if they have
fallen over toward the sensor)
• Somewhat corrected when projecting to
ground range using DEM
Kīlauea Caldera, Hawaii
Image courtesy G. Funning
• User must define an event
• Highlights the areas where the ground surface has changed
• Landslides change the backscatter intensity of the land cover by:
• Geometrical changes
• Surface property changes (e.g., vegetation, wetness)
• Changes can be positive or negative
• 𝐼𝑟𝑎𝑡𝑖𝑜 < 0 when post-event backscatter intensity increases
• To reduce noise and improve data quality we construct pre- and post-event stacks by taking
the median of all available pre- and post-event SAR imagery
Methods
SAR Backscatter Intensity Change
Log-ratio
𝐼𝑟𝑎𝑡𝑖𝑜 = 10 ∗ log10
𝐼𝑝𝑟𝑒
𝐼𝑝𝑜𝑠𝑡
𝐼𝑝𝑟𝑒 = pre-event backscatter Intensity
𝐼𝑝𝑜𝑠𝑡 = post-event backscatter Intensity
Methods
SAR Backscatter Intensity Change
Pre-event Post-Event Backscatter Change
/ =
2018 rainfall-triggered landslides Hiroshima, Japan
Images from S. Jones
Log-ratio
𝐼𝑟𝑎𝑡𝑖𝑜 = 10 ∗ log10
𝐼𝑝𝑟𝑒
𝐼𝑝𝑜𝑠𝑡
𝐼𝑝𝑟𝑒 = pre-event backscatter Intensity
𝐼𝑝𝑜𝑠𝑡 = post-event backscatter Intensity
Methods
• We compare our SAR-based
detection to an external landslide
inventory using ROC curves
• Each pixel is classified as a landslide
if its value exceeds a threshold
• For each threshold we calculate the
false positive rate and true positive
rate
• Best performance is determined by
maximizing the AUC
• AUC = 1 is perfect classifier, AUC = 0.5
is random selection
Receiver Operating Characteristic curves (ROC)
Martin Thoma - Own work -
https://en.wikipedia.org/wiki/Rec
eiver_operating_characteristic
• C-band wavelength
(5.6 cm)
• 10 m pixel for GRD
• Data collected along
ascending and
descending geometry
• Minimum revisit time 6
days for single satellite
pass
Data Collection
Copernicus Sentinel-1 A/B Satellites
Methods
• Cloud-based geospatial processing platform
• Available to scientists, researchers, and developers for analysis of the Earth's
surface
• Contains a catalogue of satellite imagery and geospatial datasets (including
Sentinel-1 SAR, Landsat and Sentinel-2 optical)
• Uses Java script and Python code editor
• Sign up for a free account https://earthengine.google.com/
• Not required to download large datasets or install specialized software!
Google Earth Engine Image from GEE
Methods
• Precipitation
• GPM, ERA5, TRMM, PRISM
Google Earth Engine
NASA GES DISC at NASA Goddard Space Flight Center
Image from Google Earth Engine
Methods
• Optical and Radar Imagery
• Sentinel, Landsat, MODIS, National Agriculture Imagery Program (US)
Google Earth Engine
European Union/ESA/Copernicus
Image from Google Earth Engine
Methods
• Terrain
• NASADEM (30 m), SRTM (30 m), USGS NED (10 m), ALOS DSM Global 30 m
Google Earth Engine
NASA / USGS / JPL-Caltech
Image from Google Earth Engine
Methods
• Lithology and Land Cover
Google Earth Engine
Conservation Science Partners
Image from Google Earth Engine
Methods
Our approach requires two simple inputs:
1) Location (Area of Interest)
2) Event timing
Additional parameters:
• Slope and curvature thresholds
• Smoothing filters
https://github.com/MongHanHuang/GEE_SAR_landslide_detection
Methods
• Provides Sentinel-1 Ground Range Detected (GRD) products
• Orthorectified
• Pre-processed to remove thermal noise
• Have undergone radiometric calibration
• New data uploaded within 2 days
• The main advantage of using GEE is that it utilizes open-access data, does not
require specialized software, analyses performed on the cloud!
• Event of Interest (EOI)
June 28, 2018 – July 8, 2018
• Record breaking rainfall event
• ~8000 triggered landslides
• 108 fatalities
• Landslide inventory provided
by the Geospatial Information
Authority of Japan (GSI) and
Association of Japanese
Geographers (AJG)
Hiroshima Prefecture, Japan
Case Studies
(Adriano et al., 2020; Hirota et al., 2019; Miura, 2019) Slide courtesy of S. Jones
The sites of landslides in Kure, Hiroshima prefecture,
southwestern Japan. Photo by Kyodo News via Associated Press.
Hiroshima Prefecture, Japan
Case Studies
• ~277 km2 AOI
• ~3800 landslides mapped in
our AOI
Next we examine data to determine the best approaches for
landslide detection and for rapid response
Handwerger et al., 2020, preprint
ROC Curves
• Combining flight paths improves
AUC score
• Addition of the DEM mask slope
and curvature thresholds
improves AUC even further
Hiroshima Prefecture, Japan
Results
Handwerger et al., 2020, preprint
Increased AUC
Results
Hiroshima Prefecture, Japan
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
Results
Hiroshima Prefecture, Japan
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
Results
Hiroshima Prefecture, Japan
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
Results
Hiroshima Prefecture, Japan
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
Results
Hiroshima Prefecture, Japan
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
Results
Handwerger et al., 2020, preprint
ROC Curves
• We explore how the total number
of images used in the pre- and
post-event stacks impacts
landslide detection
• The AUC score continues to
improve with increasing number
of SAR images (or time)
• Best approach is to use all
available images before and after
event!
Hiroshima Prefecture, Japan
Hiroshima Prefecture, Japan
Results
Iratio
1 day after event (2 SAR images)
All available images
Handwerger et al., 2020, preprint
All available pre-event and limited post-event SAR data
Hiroshima Prefecture, Japan
Results
Iratio
All available images
Landslide Heat Map
Handwerger et al., 2020, preprint
All available pre-event and limited post-event SAR data
1 day after event (2 SAR images)
Hiroshima Prefecture, Japan
Results
1 week after event (4 SAR images)
Iratio
All available images
Landslide Heat Map
Handwerger et al., 2020, preprint
All available pre-event and limited post-event SAR data
Summary
• Developed a new methodology to identify landslides (and other ground surface
changes) using open-access data
• Does not require specialized SAR processing software
• All data is stored and processed in GEE
• Best approach combined all available SAR images, ascending and descending data
with DEM mask
• Landslide density “heat maps” help reduce noise and highlight landslides for rapid
response
• Future SAR missions, like the L-band NASA-ISRO NiSAR mission, which is currently
expected to launch in January 2022, will also provide publicly available data