WEBVTT
Kind: captions
Language: en-US

00:00:01.440 --> 00:00:05.360
Hi. My name is Xinxiang Zhu.
Today I would like to talk about urban

00:00:05.360 --> 00:00:09.360
geodesy along the Hayward Fault –
automated detection of planar

00:00:09.360 --> 00:00:14.335
primitives in mobile Lidar
data and change detection.

00:00:15.931 --> 00:00:19.440
The earthquake sheared the Earth
with a unique a [inaudible] landform

00:00:19.440 --> 00:00:22.960
caused by the faulting process.
To study this process,

00:00:22.960 --> 00:00:26.560
one of the initial observations
one can acquire is measurement

00:00:26.560 --> 00:00:29.416
of fault-related
land deformation.

00:00:29.440 --> 00:00:31.280
For example,
shown in this figure,

00:00:31.280 --> 00:00:36.536
a straight fence is deformed due
to the strike-slip fault movement.

00:00:36.560 --> 00:00:41.576
Ground displacement in the far field has
an analytical form, shown on the left,

00:00:41.600 --> 00:00:46.136
where the number of the contour shows
the normalized displacement magnitude,

00:00:46.160 --> 00:00:50.776
and arrows on the dashed lines shows
the direction of the ground motion.

00:00:50.800 --> 00:00:53.920
This analytical pattern has been
confirmed by, for example,

00:00:53.920 --> 00:00:58.823
the InSAR measurement of the
2003 Bam, Iran, earthquake.

00:01:00.400 --> 00:01:04.000
In the near field, within hundreds
of meters from the fault trace,

00:01:04.000 --> 00:01:06.080
the ground displacement
is not linear.

00:01:06.080 --> 00:01:11.256
For example, the deformation
of the fence is highlighted in red.

00:01:11.280 --> 00:01:15.120
An analytical form of off-fault
displacement profile is shown on the

00:01:15.120 --> 00:01:21.520
left, where the X axis represent distances
from a fault trace, and Y axis represent

00:01:21.520 --> 00:01:26.216
ground displacement induced by
various underlying rock properties.

00:01:27.200 --> 00:01:30.640
This profile only exhibit
distinguishable curvatures

00:01:30.640 --> 00:01:35.896
in the near field from which we
can study the shallow-slip deficit.

00:01:35.920 --> 00:01:41.336
However, the corresponding geodetic
signature is extremely hard to collect.

00:01:41.360 --> 00:01:44.320
This is because the curvature
require dense observations

00:01:44.320 --> 00:01:48.320
in the near field to quantify,
therefore limit the application

00:01:48.320 --> 00:01:52.056
of traditional sparse
geodetic observations.

00:01:52.080 --> 00:01:54.720
And the scale of the change
can be smaller than the

00:01:54.720 --> 00:01:59.119
critical detection threshold
of some geodetic tools.

00:01:59.920 --> 00:02:02.720
In this study, we propose
a new method using

00:02:02.720 --> 00:02:08.254
mobile Lidar data to detect fault
displacement in the near field.

00:02:09.280 --> 00:02:13.520
Using mobile Lidar data sets,
we explore the Lidar change detection

00:02:13.520 --> 00:02:17.280
strategies and develop this
method called modeling-based

00:02:17.280 --> 00:02:21.840
change detection using
geometric primitives.

00:02:21.840 --> 00:02:26.240
Shown here are our change detection
result for the 2014 South Napa

00:02:26.240 --> 00:02:29.600
earthquake where point cloud
are model are simple

00:02:29.600 --> 00:02:34.320
geometric models that can be
described by a few free parameters,

00:02:34.320 --> 00:02:37.687
so-called geometric primitives.

00:02:38.400 --> 00:02:43.360
By model the 1-yard rows as planar
primitives, we were able to recover

00:02:43.360 --> 00:02:46.696
25-centimeter
coseismic displacement,

00:02:46.720 --> 00:02:52.743
which agree with in situ and alignment
array measurement at decimeter level.

00:02:53.360 --> 00:02:58.560
By model the anchor posts as cylindrical
primitives, we were able to recover the

00:02:58.560 --> 00:03:04.696
postseismic displacement approximately
one month after the main shock.

00:03:04.720 --> 00:03:07.760
And the detected postseismic
displacement agree with

00:03:07.760 --> 00:03:13.272
co-located alignment array
measurements at sub-centimeter scale.

00:03:14.320 --> 00:03:18.160
Here shows the postseismic
displacement profile revealed by

00:03:18.160 --> 00:03:23.200
the cylindrical primitives,
where the fault-parallel displacement

00:03:23.200 --> 00:03:29.044
and the angular change are distributed
against off-fault distances.

00:03:30.800 --> 00:03:34.640
The reason that we achieve
high accuracy change detection

00:03:34.640 --> 00:03:38.776
is because we model
point cloud as primitives.

00:03:38.800 --> 00:03:43.600
Shown in this figure, the reference and
the secondary point cloud can be treated

00:03:43.600 --> 00:03:49.656
as Lidar measurements collected
before and after a fault movement.

00:03:49.680 --> 00:03:54.960
And, if we model point cloud as plane
surface and calculate the point-to-point

00:03:54.960 --> 00:04:01.200
distances, the model-to-model distances
are a more robust estimator of the

00:04:01.200 --> 00:04:07.209
true displacement compared
with point-to-point distances.

00:04:07.920 --> 00:04:13.200
This is because the point-to-point
correspondence is highly unstable

00:04:13.200 --> 00:04:17.394
due to the irregular
format of the point cloud.

00:04:18.560 --> 00:04:23.976
Here we show experiment where a
displacement is randomly generated,

00:04:24.000 --> 00:04:26.560
and the two change
detection strategies

00:04:26.560 --> 00:04:31.450
are performed to recover
the simulated deformation.

00:04:32.320 --> 00:04:36.536
The residual of the recovered
deformation is found here.

00:04:36.560 --> 00:04:41.280
And the modeling-based change
detection result is smallest residual

00:04:41.280 --> 00:04:46.296
compared with change detection
using point-to-point distances,

00:04:46.320 --> 00:04:50.400
which is the default metric
behind a popular change detection

00:04:50.400 --> 00:04:56.377
algorithm called iterative
closest point, or ICP.

00:04:57.360 --> 00:05:02.000
ICP is the dominant method
processing Lidar point cloud

00:05:02.000 --> 00:05:05.656
for the purpose of fault
deformation monitoring.

00:05:05.680 --> 00:05:09.840
And, as you can see, the modeling-based
change detection outperformed the

00:05:09.840 --> 00:05:15.527
ICP method by almost
one order in this experiment.

00:05:16.160 --> 00:05:21.360
Therefore, the modeling-based change
detection using geometric primitives

00:05:21.360 --> 00:05:26.400
is preferred to detect subtle fault
creep displacement that is

00:05:26.400 --> 00:05:32.134
characterized by steady and
gradual deformation over years.

00:05:33.680 --> 00:05:38.080
In order to model primitives in the
urban environment, we adapted a

00:05:38.080 --> 00:05:43.896
detector based on the classic random
sample consensus, or RANSAC.

00:05:43.920 --> 00:05:47.840
This new detector can extract
corresponding planar primitives

00:05:47.840 --> 00:05:51.840
parallelly from the reference
and the secondary data sets

00:05:51.840 --> 00:05:55.656
collected before and
after the change.

00:05:55.680 --> 00:06:01.496
For example, planar surface, like
building walls, roofs, and garage doors

00:06:01.520 --> 00:06:07.896
can be extracted parallelly by the detector
to trace the fault creep deformation.

00:06:07.920 --> 00:06:12.720
And the corresponding primitives
are color-coded and ordered by

00:06:12.720 --> 00:06:17.110
the number of points
exist in both data sets.

00:06:19.040 --> 00:06:24.240
Using the corresponding planar
primitives, a shear deformation

00:06:24.240 --> 00:06:28.216
represent fault creep
displacement can be estimated

00:06:28.240 --> 00:06:31.816
from augmentation
of primitives.

00:06:31.840 --> 00:06:35.680
In other words, the fault creep
displacement of a house,

00:06:35.680 --> 00:06:39.280
shown in the previous slide,
can be estimated by tracking

00:06:39.280 --> 00:06:43.600
the coincident movement of these
walls, roofs, garage doors,

00:06:43.600 --> 00:06:48.758
and other planar surfaces that are
extracted by the new detector.

00:06:49.680 --> 00:06:53.840
This estimation uses
a combined least square regression

00:06:53.840 --> 00:06:58.756
that is originally developed
for Lidar sensor calibration.

00:07:01.520 --> 00:07:06.160
Using this new method, we analyzed
mobile Lidar data collected along

00:07:06.160 --> 00:07:10.880
a 2-kilometer segment of the
Hayward Fault, and we validate

00:07:10.880 --> 00:07:14.720
the change detection result
with fault creep measurements

00:07:14.720 --> 00:07:19.496
of an alignment array located
at Camellia Drive.

00:07:19.520 --> 00:07:23.520
Here shows the 10-year fault
creep displacement detected by

00:07:23.520 --> 00:07:28.430
satellite surveys at the
alignment array stations.

00:07:30.560 --> 00:07:35.256
Again, here shows our study area
and our change detection result.

00:07:35.280 --> 00:07:39.360
For visualization purposes,
let’s down-sample our displacement

00:07:39.360 --> 00:07:44.320
vector a little bit and
give it a imagery base layer.

00:07:44.320 --> 00:07:47.440
Here shows the change
detection result for the

00:07:47.440 --> 00:07:54.216
Hayward Fault creep observed
from July 2015 to June 2017.

00:07:54.240 --> 00:07:58.376
The arrow represent the
ground displacement vector

00:07:58.400 --> 00:08:02.856
with the size proportional
to the change.

00:08:02.880 --> 00:08:07.600
If we project the displacement vector
along the fault trace, we can get this

00:08:07.600 --> 00:08:13.656
heat map where the color shows the
normalized fault-parallel displacement.

00:08:14.560 --> 00:08:18.560
From the change detection result,
we find uniform strike-slip

00:08:18.560 --> 00:08:22.960
displacement in the far field.
And, in the near field,

00:08:22.960 --> 00:08:27.040
we find this interesting
rotationing pattern that are

00:08:27.040 --> 00:08:31.367
rarely observed from
other geodetic data sets.

00:08:33.440 --> 00:08:38.640
To validate our change detection result,
we compare the detected fault creep

00:08:38.640 --> 00:08:44.776
displacement with co-located alignment
array measurements at Camellia Drive.

00:08:44.800 --> 00:08:48.800
Arrows shows in this plot
represent fault creep displacement

00:08:48.800 --> 00:08:54.320
detected by our method.
And average fault-parallel displacement

00:08:54.320 --> 00:08:59.758
near the alignment array benchmarks
are calculated for validation.

00:09:00.800 --> 00:09:04.240
The comparison results are
reported here where the difference

00:09:04.240 --> 00:09:09.176
between the two measurements
is at sub-centimeter scale.

00:09:09.200 --> 00:09:15.037
And it is also within the detection
uncertainty of both data sets.

00:09:16.400 --> 00:09:22.000
Referring to the fault trace profile of
fault-parallel and fault-perpendicular

00:09:22.000 --> 00:09:28.226
displacement can be generated and
shown in the middle and bottom plot.

00:09:28.960 --> 00:09:33.440
Alignment array measurements almost
overlap with the detected fault-parallel

00:09:33.440 --> 00:09:39.302
displacement, which validate our
change detection result at this location.

00:09:41.200 --> 00:09:46.480
It also appears that benchmark IS
is located within the nonlinear

00:09:46.480 --> 00:09:52.480
deformation zone such that the baseline
of alignment array does not span the

00:09:52.480 --> 00:09:58.080
entire near-field deformation zone,
which indicated by the angular change

00:09:58.080 --> 00:10:02.290
of the displacement vector
across the fault trace.

00:10:03.200 --> 00:10:09.656
This suggests that the ES-IS baseline
does not span the entire creeping zone,

00:10:09.680 --> 00:10:14.640
and the measurement of alignment
array at this location and possibly

00:10:14.640 --> 00:10:18.960
at other locations is under the
risk of underestimating the

00:10:18.960 --> 00:10:23.401
fault creep displacement
detected in the near field.

00:10:25.360 --> 00:10:31.200
To confirm that this rotational pattern
and the nonlinear deformation zone are

00:10:31.200 --> 00:10:36.640
not an artifact caused by the smoothing
effect employed by the moving

00:10:36.640 --> 00:10:41.998
window detection,
a synthetic test was conducted.

00:10:43.520 --> 00:10:48.216
In the synthetic test, at the same
location of Camellia Drive,

00:10:48.240 --> 00:10:52.640
we simulated a 4-centimeter
strike-slip displacement that

00:10:52.640 --> 00:10:58.480
matches the scale of real fault creep
with instant transition of slip amount

00:10:58.480 --> 00:11:01.381
and direction
at the fault trace.

00:11:02.320 --> 00:11:06.480
Shown in this plot are the
change detection response

00:11:06.480 --> 00:11:09.465
of the simulated deformation.

00:11:10.400 --> 00:11:14.400
This experiment represent a
conservative estimation of the

00:11:14.400 --> 00:11:19.600
smoothing effect where elastic fault
displacement pattern at the

00:11:19.600 --> 00:11:25.120
ground surface is assumed.
In reality, a gradual transition

00:11:25.120 --> 00:11:30.560
is expected due to plastic deformation
or the buried slip front,

00:11:30.560 --> 00:11:35.707
which could cause less smoothing
effect within the searching window.

00:11:37.200 --> 00:11:42.856
There is no bias found in the directions
of the detected displacement.

00:11:42.880 --> 00:11:46.160
And the angular variation
of the detected result is

00:11:46.160 --> 00:11:49.493
about 10 degrees
for 1 sigma.

00:11:51.040 --> 00:11:55.280
The synthetic result confirms that
the moving window detection

00:11:55.280 --> 00:12:00.536
does introduce a smoothing
artifact into the change detection.

00:12:00.560 --> 00:12:06.136
However, the smoothing effect
should be no larger than 10 meters.

00:12:06.160 --> 00:12:11.440
Therefore, the nonlinear deformation
detected in the real data sets shown

00:12:11.440 --> 00:12:16.376
in the previous slide is unlikey
induced by a detection artifact.

00:12:16.400 --> 00:12:20.640
Whereas, a nonlinear deformation
zone detected in real data set is

00:12:20.640 --> 00:12:26.960
at least 50 meters wide with
noticeable systematic rotation pattern

00:12:26.960 --> 00:12:31.305
that are not observed
in the simulation.

00:12:34.560 --> 00:12:39.360
Here shows the overall off-fault
displacement profile where the

00:12:39.360 --> 00:12:43.440
fault-parallel and fault-perpendicular
ground displacement are distributed

00:12:43.440 --> 00:12:46.275
versus off-fault
distance.

00:12:46.800 --> 00:12:52.271
A 2.5-centimeter dextral displacement
is detected in the far field.

00:12:53.680 --> 00:12:57.280
On the right shows the most recent
UAVSAR change detection result

00:12:57.280 --> 00:13:01.072
provided by Eric Fielding
from NASA JPL.

00:13:02.000 --> 00:13:05.760
As you can see, the two measurements
agree well in terms of the detected

00:13:05.760 --> 00:13:09.621
fault-parallel displacement
in the near view.

00:13:12.407 --> 00:13:16.320
In conclusion, in this study,
we have shown that geometric

00:13:16.320 --> 00:13:20.720
primitive model from mobile Lidar
point clouds is not only an effective

00:13:20.720 --> 00:13:25.280
feature to track the change but also
can be combined as persistent

00:13:25.280 --> 00:13:30.536
geodetic markers to improve the
change detection performance.

00:13:30.560 --> 00:13:33.440
Our new method detects
centimeter-level change with

00:13:33.440 --> 00:13:38.800
sub-centimeter level accuracy,
which enable the detection of aseismic

00:13:38.800 --> 00:13:43.491
fault creep ground displacement
over two years of observations.

00:13:44.400 --> 00:13:47.840
Validation with alignment array
confirms the uncertainty of the

00:13:47.840 --> 00:13:52.320
detection and suggests that the
baseline of an alignment array

00:13:52.320 --> 00:13:56.856
might not be long enough to
span the entire creeping zone.

00:13:56.880 --> 00:14:01.840
For more information, please
refer to our two publications.

00:14:04.240 --> 00:14:08.560
Finally, I would like to acknowledge
my co-authors, Craig Glennie,

00:14:08.560 --> 00:14:13.120
who is also my Ph.D. adviser;
Ben Brooks and Todd Ericksen

00:14:13.120 --> 00:14:18.080
from USGS; and we would also
like to thank Eric Fielding from

00:14:18.080 --> 00:14:23.176
NASA JPL for providing the
UAVSAR change detection result.

00:14:23.200 --> 00:14:25.120
Thank you.