PubTalk 2/2020 — From California to Cambodia

Video Transcript
Download Video
Right-click and save to download

Detailed Description

From California to Cambodia - Surface Water Mapping Using Cloud-Based Remote Sensing (by Christopher Soulard, USGS Research Geographer)

  • With cloud based remote sensing we can now access, process, and analyze terabytes of satellite imagery remotely via the internet.
  • Satellite images and streamgages are the two most common measurement tools used to assemble a record of historical surface water changes.
  • Weekly to monthly snapshot of surface water from satellite images can identify the extent of discrete flooding events.
  • These maps, along with climate records, enable us to better understand how precipitation events contribute to the timing, location, and extent of flooding and drought.

Details

Date Taken:

Length: 00:56:08

Location Taken: Menlo Park, CA, US

Transcript

0:00
Well, I’m Susan Benjamin. I’m with the U.S. Geological Survey,
0:04
and I’d like to welcome you all to the USGS monthly public lecture.
0:11
And one thing they want me to do is fill you in about next month’s lecture
0:16
so that you can all come back then. Next month’s lecture –
0:21
there’s a flier out for it. It’ll be on March 26th, 7:00 p.m.
0:27
And it’s a talk by Sara McBride, who is a USGS research social scientist,
0:33
on social science and natural disasters.
0:36
So hopefully your calendars are free, and you guys can come back for that.
0:41
A little – a little bit of housekeeping.
0:44
If you need to use the restroom, it’s out the back door and around to the right.
0:49
In the unlikely event that the fire alarms go off, we will all walk calmly and
0:56
quietly out the back of the auditorium and down the stairs and meet up
1:00
in front of the building. But that’s never happened yet.
1:04
So don’t hold your breath. And finally, we hope that you
1:08
can hold your questions during the lecture to the end.
1:11
And then, because we’re recording this, we’d like for you
1:14
to use the microphones right there to ask questions.
1:18
So I am delighted to introduce our speaker today, research geographer
1:23
Chris Soulard from the USGS Western Geographic Science Center.
1:27
Chris holds two bachelor’s degrees from UC-Santa Barbara
1:30
as well as a master’s degree from San Jose State.
1:33
And he has been with USGS for the past 18 years.
1:35
His academic and professional careers have centered around characterization
1:39
of the Earth’s surface using satellite imagery, aerial photography,
1:43
and terrestrial Lidar instrumentation.
1:47
Most of Chris’ career was dedicated to the Land Cover Trends Project,
1:50
which used the decades-long record of Landsat satellite imagery
1:54
to track landscape changes throughout the nation.
1:57
He led many facets of this research in the western U.S., including
2:00
writing over 10 ecoregion reports that summarized regional trends
2:04
in land use as mapped by the project team.
2:07
Chris instigated publication of all the field photos collected over
2:11
a decade of field campaigns for that project, which includes
2:14
about 30,000 geo-tagged photos now available to the public.
2:19
Chris has written more than 50 peer-reviewed reports over
2:22
his career and is now the principal investigator for the PLACE Project,
2:26
which is responsible for large-scale assessments of historic change and
2:30
driving forces to better understand how landscapes may change in the future.
2:35
And he’s going to be talking today about
2:36
mapping surface water using cloud computing.
2:40
So thank you. - Thank you.
2:42
[Applause]
2:46
Hope everyone – can everyone hear me okay?
2:49
So, as Susan mentioned – thank you for the introduction, Susan –
2:52
my name is Chris Soulard. I’m with the USGS Western
2:54
Geographic Science Center. And about six months ago, I lived and
2:57
worked about 100 feet away from here. We’re now down in Moffett Field.
3:02
So, while the location may have changed, the work has
3:04
remained the same. And so today I’m here to talk about
3:08
the core part of my project, where we’re focusing on multi-decade
3:13
collections of satellite imagery, which we’re using to compile
3:16
monthly records of surface water. In California in that image to the left.
3:22
In Cambodia – image to the right. And now we’re able to
3:24
do that readily anywhere on the planet.
3:28
As Susan mentioned, I’m the principal investigator on the PLACE Project.
3:31
And what that stands for is Patterns in the Landscape –
3:34
Analyses of Cause and Effect.
3:36
We’re part of the third iteration of the Land Cover Trends project,
3:40
which Susan also mentioned was responsible for mapping land use and
3:44
land cover change from 1973 to 2000, and in some cases, to 2010,
3:50
across satellite images. And those were collected in
3:53
five- to eight-year increments. And, just to give you a little bit of
3:57
background on land use land cover mapping, it’s a pretty simple concept.
4:00
Essentially, we take an image, and we color it up.
4:03
Where every pixel is colored up based on the predominant land use
4:07
or land cover that makes up the landscape at that given point in time.
4:11
So that might be a characterization of urban landscapes, of agricultural lands,
4:15
of forest or shrubland. It could be even specific types of forests.
4:18
But ultimately, we draw that classification, and by doing that
4:22
sequentially, we can look at how classes change from
4:24
one class to another over time. Compile these – and then we can
4:28
summarize those changes and say, okay, the Pacific Northwest has
4:31
a long history of logging. We can actually report those rates of logging.
4:34
We can reports rates of development in California or other parts of the country
4:38
that are maybe growing pretty rapidly. We’ve learned quite a bit from that
4:42
assessment, and I actually have some reports in the back there,
4:44
if you’re interested, that summarize all those changes
4:46
in the western United States from 1972 to 2000.
4:50
This happens to be an example of that work from Central Valley where we –
4:54
the orange color represents agricultural lands.
4:57
The red color represents developed lands.
4:59
And so, in this particular sample, we see agricultural lands being lost
5:03
at the expense of development in that particular sample plot.
5:08
So, as I mentioned, this work was done in five- to eight-year increments.
5:11
The PLACE Project is changing the game, in that we’re now focusing on
5:16
studying changes on annual – in annual windows, so we can
5:20
understand short-term change events, and then sometimes sub-annual events.
5:24
So, as we’ll talk about with these – the surface water maps,
5:26
we’re doing so on a monthly scale.
5:28
And once we have those records in hand, we can start to understand
5:32
what caused those. Why are those changes taking place,
5:35
when they’re taking place, to try to better understand how to
5:39
potentially mitigate against negative impacts in the future.
5:43
Here’s another way of visualizing how our project operates.
5:46
This is a little headier, but I’ll walk you through it.
5:49
This happens to be a stack of satellite images – 15.
5:52
We’re now looking at about 300 or so images at a time.
5:56
And so, at every single image, at every single pixel in that image – and just to –
6:00
an aerial – an aerial photo or a satellite image is no different from a picture
6:04
that you collect on your iPhone. It’s actually – if you zoom in on
6:07
those pictures, you see that it starts to get grainy. Those are pixels.
6:10
All those individual pixels, we can take discrete measurements.
6:13
We can characterize land use or land cover type.
6:16
We can also characterize vegetation greenness with satellite imagery and
6:20
track that progression over time. So it could be an investigation
6:24
of phenology. It could be water stress on a plant.
6:28
Those are the types of things we can investigate.
6:30
By compiling those records over time, then we can say okay, well,
6:34
why are we seeing more change in this location versus another?
6:37
Why are we seeing that this period is showing a higher rate of change,
6:42
whereas the period that follows it is a much slower rate of change?
6:46
So, to give some real-world examples here, we think about development.
6:51
We think about California’s historic rates of growth of development.
6:56
We have a – we could probably all remember points in California
6:59
where there were real boom cycles of development.
7:01
There’s also times like the Great Recession, where we had
7:04
real bust cycles, right? Where we just – where the
7:06
rate of development slowed. So we can actually chronicle
7:08
that story and then pinpoint that the Great Recession may
7:13
have triggered that slowdown. And that’s just one example of many.
7:17
As for the locations, we can look at different communities at one time.
7:20
So if we’re focusing, again, in the western United States,
7:23
we might question, well, why does Las Vegas see rapid rates of growth,
7:27
whereas, Oakland, California, doesn’t? This is a obvious answer.
7:30
It’s a softball where, you know, there’s only so much space where
7:33
you can develop in Oakland, and there’s probably some sort of
7:36
growth policies in place as well. Vegas, on the other hand,
7:39
lots of abundant land. There’s probably good reasons – there’s lots
7:41
of reasons why you can see faster rate of growth in Las Vegas versus Oakland.
7:45
So drastic example, but I hope that helps you understand a little bit more of
7:51
what we can do once we have these time series of change at our fingertips.
7:57
So this paradigm applies to lots of different land use and land cover types.
8:02
Over the last couple years, we’re focusing specifically
8:04
on surface water and surface water change over time.
8:08
You may ask, why surface water? Well, three years ago, I wrote a paper
8:11
with Christine Albano and Mike Dettinger, where we look at –
8:14
looked at atmospheric river events and atmospheric rivers, which I’m sure
8:18
you’re all aware of from weather reports, are these large moisture-bearing
8:22
tropical storms from the Pacific that oftentimes bring upwards of 80% of
8:27
our precipitation in California and to the western U.S. as a whole.
8:31
And so what we did in that paper is, we took atmospheric river events
8:34
and identified all the precipitation that fell during those events and linked
8:39
that to vegetation green-up and subsequent fire risk in the western U.S.
8:43
On the heels of that project, we asked ourselves, well, what else
8:46
are atmospheric rivers influencing? And thinking back to those weather
8:50
reports, every winter – you could all probably say it with me – we hear,
8:54
atmospheric river coming to the Bay Area.
8:57
Flooding expected in these communities.
8:59
And so we know that the Bay Area is affected by atmospheric rivers,
9:03
but actually, an area that’s affected even more than the Bay Area
9:05
is the Central Valley. The Tulare Basin by Visalia is an area
9:11
that’s especially affected by floods, historically speaking.
9:14
And this is a rendition of the Great Flood of 1862 that led to
9:17
inundation in the state’s capital for, I think, three weeks, if not more.
9:23
And so these large events – historic events are what led us
9:25
to focus on the Central Valley.
9:29
Also, our background in remote sensing and our experience mapping surface
9:32
water over these five- to eight-year intervals led us to believe that we could
9:35
confidentiality map surface water over denser time series pretty easily.
9:41
And I say “easily,” but it wasn’t as easy
9:43
as we initially thought it was going to be.
9:46
And, of course, this work, we knew, would immediately feed into and
9:51
fulfill USGS mission goals by identifying risks –
9:54
being identifying actual flood events.
9:57
And then by sharing that information with the public and with land managers,
10:01
we could help mitigate against those risks.
10:04
So here’s a simplified analogy to explain how we go about our work.
10:11
And I had some fancy fonts that made it look a little more mysterious,
10:14
but I call this the Mystery of the Kitchen Puddle.
10:17
So let’s all imagine, we wake up one morning, and we have a puddle
10:19
in our kitchen. What’s the first question you ask yourself?
10:23
Probably, why the heck is this puddle here, right?
10:26
[laughs]
10:27
And so you start to go through potential reasons why this may be occurring.
10:31
You may say, okay, maybe there’s a leak in my dishwasher.
10:34
Maybe my spouse or my dog or my – or my child spilled a big cup of water.
10:40
Or maybe it’s somehow associated with the rainstorm event
10:45
actually raining on my community right now.
10:48
But, before you can actually say what the cause of that is, you have to
10:52
start to compile evidence. So you may go check that pipe.
10:55
You may ask around your house, and hopefully you have reliable –
10:58
you have good – your good interrogation skills.
11:03
Or you might just think back to past events that have affected your home.
11:08
So let’s just imagine – this is a hypothetical, of course,
11:10
but let’s just say you think back to the winter of 2017,
11:13
and you had another puddle in 2017. It was a little smaller.
11:17
And you did ask your kids, who were babbling at the time because they
11:21
couldn’t speak – they couldn’t speak yet. And you – and you checked
11:25
the pipe in that case. And so it wasn’t the pipe then,
11:28
and it wasn’t your – and it wasn’t your kids, and it wasn’t your spouse.
11:31
And so you figure, okay, well, maybe – is this somehow related?
11:34
And then you think back maybe five years prior to that, and you had another
11:37
big puddle in your upstairs room. And there wasn’t any pipes up there.
11:42
And you didn’t have kids yet. So these past records can help inform
11:46
what may have happened in this current event that you’re trying to understand.
11:50
So let’s – these are just two – you know, two potential examples
11:53
that you could reach back your mind and access and use as evidence.
11:57
But you may have 10 events to pull from.
11:59
And, by having sequential events that you could start to tease out what did
12:03
and did not contribute to those events, you might be able to
12:05
identify the culprit. You may say, wait, all these
12:07
are happening kind of wintertime or springtime when rainfall occurred.
12:11
So maybe these are actually due to that rainfall event.
12:15
Well, this is exactly what we’re trying to do in California.
12:18
Every winter, puddles arise in California.
12:20
You can call them floods. I’ll call them puddles.
12:23
And they’re different sizes, different locations,
12:25
they cluster in certain pockets of the state.
12:28
But, before we can say why these puddles are occurring when they occur,
12:31
we need to compile evidence.
12:33
And this is where the satellite record comes into play.
12:36
Like, mapping all these flood events every single year over that time –
12:40
over – sequentially, year or year, or month over month, you can –
12:45
you can amass a comprehensive record, or, in an ideal situation,
12:50
amass this comprehensive record of floods for which
12:53
to then compare to possible explanations.
12:57
Maybe floods are occurring where you have the heaviest precipitation.
13:00
Maybe it’s areas that have a high rate of snow melt.
13:03
Maybe it’s areas that have poor soil saturation capability.
13:07
Maybe it’s areas that have a lot of impervious surfaces.
13:10
All those have been reported in literature to be contributing factors.
13:13
And so, once we have that record, we can go compare those maps
13:17
to those possible explanations through different regression techniques
13:21
to try to identify why the floods occurred when they did.
13:26
As remote sensors, we have a pretty expansive toolkit to choose our tools
13:31
from and to compile that evidence and to measure the Earth’s surface.
13:36
All these satellites – a lot of them I don’t really know personally –
13:41
have different strengths and different weaknesses.
13:45
And the ones I circle here are the one that I have worked with exclusively –
13:48
or, primarily in my career. And I’ll talk about how we
13:51
considered these as possible tools to investigate surface water.
13:56
The Landsat program got started in 1972 and has been going
13:59
strong through today. So we have 50 years of records,
14:04
essentially, from the Landsat archive. And this is a measurement of
14:08
every location on Earth every 16 days for that period of time.
14:14
So you have 26 – upwards of 26 observations a year where you
14:17
could potentially look at phenomena, both statically and then over time.
14:23
It is a reduced pixel resolution, so this is what a Landsat image
14:28
looks like over a baseball stadium. So it’s fairly pixelated.
14:30
It’s kind of hard to make out a lot of detail of that baseball stadium.
14:34
But just imagine that temporal richness and what you could
14:36
really tell over time. You could actually observe
14:39
baseball stadiums being built with Landsat satellites.
14:43
It’s an optical multi-spectral scanner. And that doesn’t mean anything to you
14:47
right now, but it will mean something when I talk about what it can and what
14:50
it cannot do when it comes to detecting surface water very shortly.
14:54
And I’ll mention, while the project – while the program got started in 1972,
14:58
1982 is when the 30-meter satellite launched.
15:01
And so that’s where we have 40 years of records that have – where we have
15:04
30-meter resolution data continuously moving forward to today.
15:11
So the MODIS satellite is similar to Landsat, and that is also
15:15
optical multi-spectral scanner. And just keep that in the
15:18
back of your mind. There will be relevance soon enough.
15:21
It’s a coarser spatial resolution, so rather than 30-meter pixels,
15:25
you’re talking about 250- to 500-meter pixels.
15:28
So that baseball stadium is really just one or two pixels,
15:31
so very coarse spatially. And it’s only been around for two decades.
15:37
But something that will be important is that it has a temporal resolution
15:40
of every day. So, rather than having snapshots every 16 days,
15:45
we now have a continuous look at the – we can look at the landscape every day.
15:49
So this is really good for understanding short-term change events –
15:53
like floods, perhaps. And we’ll get to that. But also phenology.
15:56
This is – MODIS is readily used in a lot of phenological studies when you want
16:00
to understand green-up trends and die-back and exact – the exact timing
16:03
of that and how it changes over time, potentially due to climate changes.
16:07
But that – so, but I regress for now and will – or, digress for now,
16:12
and will get back to MODIS shortly.
16:15
And then Sentinel is party of a different family of satellites
16:17
all together called radar satellites.
16:20
It’s important to note these because radar satellites offer
16:23
something that optical sensors don’t.
16:26
The active radar pulse that’s used in these satellites actually allow you
16:30
to penetrate clouds. And clouds bring the moisture that lead to floods.
16:34
So it would be really nice to see through the clouds when we’re trying
16:37
to represent the Earth’s surface when we think a flood is occurring.
16:42
A major drawback to Sentinel is that it’s only been around for six years.
16:46
So great for surface water detection in cloudy times, but pretty short-lived.
16:55
So our question to ourselves was, what tool do we choose to move
16:59
forward and try to investigate surface water?
17:02
And so, as I mentioned, do we take something that has no cloud problems,
17:06
but short-lived, or do we take something that has great temporal
17:09
resolution but poor spatial resolution and maybe has only been around
17:12
for two decades? Or do we take something that has
17:14
a very expansive record and may have a only 16-day revisit rate,
17:19
but at least it’s been around for 40 years?
17:21
Well, for us, because our core goal is to tie this back to long-term
17:24
precipitation records, we decided to use Landsat initially as our method
17:30
to go ahead and try to map the Earth’s surface.
17:33
And this just shows how problematic clouds can be.
17:36
This is a MODIS composite. And clouds were a problem.
17:39
And I will eat that horse [laughs] as we go through this presentation.
17:45
So, as with any other research the USGS conducts, our due diligence is to do a
17:50
lit review before we get going to find out what’s been done in the space.
17:54
And we quickly realized there’s been a lot of work from some other
17:57
leaders in this space. Dartmouth Flood Observatory has done work
18:01
with MODIS, Sentinel, Landsat – they’ve done almost everything,
18:05
and we learned a lot – got to leapfrog a lot over the learning curve
18:10
by reading up on their work.
18:11
The European Commission Joint Research Center collaborated with
18:14
Google a few years ago and actually created a monthly water presence and
18:19
absence product. Sounds familiar. So that’s what we’re trying to get,
18:21
right, using Landsat imagery. So I highlight that one because
18:26
that’s something that we ultimately used for our original research when
18:30
we realized that they had done some of the heavy lifting for us.
18:34
University of Maryland has done some work in this space that’s worth noting,
18:37
especially on radar recently. And then the USGS has been
18:40
doing some work here for some time.
18:41
Dr. John Jones developed a Dynamic Surface Water Extent model.
18:46
And that’s a model that maps different levels of inundation in Landsat imagery.
18:51
So we highlight that one because we actually are going to use
18:54
that one moving forward as well. So those two are the two products
18:56
that we decided to use in our early exploratory efforts.
19:03
If we had started this project five years ago, we would have – we probably
19:07
wouldn’t be reporting to you today. It would be – it’s a big lift to try to
19:11
map any land use or land cover phenomena continuously across
19:14
the satellite archive. So, if we were – had done that,
19:18
we would download gigabytes of data to a local machine like this,
19:21
pre-process that, use expensive software, and maybe after a year,
19:25
we would have some results to share. Well, the last five years has
19:28
changed things dramatically. So cloud computing platforms,
19:33
like Google Earth Engine, now allow us to access our own imagery within –
19:38
on Google servers. This happens to be a server in Iowa, I believe, where
19:42
they have millions of our images stored. And we can go ahead and send –
19:47
write some code, send in a request to process those images, analyze those
19:52
images, and shoot us back results. And that – on the order of one day,
19:57
we can have the same results that would have taken us a whole year
19:59
to do just five years ago. So it’s pretty amazing,
20:02
and we heavily leverage Google Earth Engine for our research.
20:05
Otherwise, we wouldn’t be able to do big data analysis
20:07
like we’re currently doing.
20:09
So here’s the workflow that we decided to employ
20:13
in our first of our published papers.
20:16
We took JRC maps that were available in Google Earth Engine directly, which,
20:21
again, are water presence and absence maps produced monthly.
20:25
We decided to code up the Dynamic Surface Water Extent model within
20:29
Google Earth Engine to create a second monthly product.
20:33
And so now we had two products created monthly from 1985 to 2015,
20:38
so 372 months of data that we could go ahead and play with.
20:45
But we wanted to validate those maps and figure out if they were good or not.
20:48
So the first thing we had to do was to find a source to validate against.
20:51
Well, USGS has a very robust stream gauge network which
20:55
you may or may not have heard of. It’s probably one of our greatest assets
20:59
as an agency. It’s made up of roughly 10,000 different gauges
21:03
that have been collecting data, in many cases, for multiple decades.
21:06
So we figured, well, let’s just see what kind of gauge records
21:09
we have for the Central Valley.
21:11
And we had a pretty robust network that we could go ahead and compare
21:14
back discharge records collected monthly to our monthly maps.
21:19
By establishing relationships between the discharge and our monthly maps,
21:22
we could go ahead and interpolate gaps based on those discharges
21:27
where we have clouds that obstruct our ability to see the landscape.
21:30
And let’s just imagine. We have January. It’s cloudy.
21:33
We can’t even see what’s happening on the landscape with our satellite.
21:36
Based on that long-term relationship, we could say, okay, well,
21:39
discharge was high that month. And we can actually interpolate
21:42
what surface area may have looked like in that month that
21:45
we couldn’t see the landscape. Now, once we have a continuous
21:48
time series, we can go ahead and do trend analysis, comparisons
21:53
to precipitation records, or even that hindcasting where we can
22:00
actually back-cast and kind of simulate what the
22:03
1862 event may have actually looked like.
22:08
So, before we can do any of that fun – the end analysis or the – all that stuff
22:13
I just mentioned at the very end of the last slide, we had to go ahead and see if
22:17
our maps were being produced properly. So we had to compare our outputs to –
22:21
produced in the cloud to USGS official outputs.
22:25
So these three maps were recently made publicly available on what’s
22:28
called Earth Explorer, which is a data repository that anyone can access.
22:33
And we found, upon comparing our data to those data, that our maps
22:37
were between 91 and 99% in agreement with those – with those maps.
22:43
We also compared the JRC maps with the DSWE map to if they were
22:47
similar to one another, and we found that they had a 0.75 correlation or
22:52
greater for most of the watersheds within the Central Valley.
22:57
The third kind of thing we – is always required is we had to go ahead and look
23:00
at the map to see what they look like. So when we – when we pulled up the
23:03
maps, and we saw that they – actually, the boundaries, the water features,
23:06
aligned with really well with the imagery when we have clear –
23:09
a clear view of the landscape, but that’s when we have
23:12
a clear view of the landscape. Now, if we had those cloud problems
23:15
that I’ve mentioned, and I will continue to mention, we have – that’s where we
23:20
omit a lot of surface water information.
23:22
And we omit a lot of flood events over those winter months, which
23:26
really hurts us when we try to compile that continuous record.
23:30
But here, just for – let’s look at a month where we don’t typically have too much
23:35
cloud obstruction, so spring in California, where we don’t –
23:39
where the – it’s usually less rainfall. This happens to be Friant Dam
23:43
just outside of Fresno. Just a few roughly five-year increments
23:47
of JRC maps just to show you that the years that we can all recall were drought
23:53
years in California coincide with the low water levels at Friant Dam.
23:56
And the years that we probably can tie back to wetter-than-normal years
24:01
have a high water level at Friant Dam. So at least Friant Dam was indicative
24:04
in our maps of Friant Dam are a – pretty good representations of
24:08
overall precipitation in that – in that time in California’s history.
24:14
So that was kind of an informal validation
24:16
that we were at least on the right track.
24:19
So going back to compiling those relationships between the stream gauge
24:23
records and our maps, we went ahead and took those discharge records that
24:27
were compiled for five years or longer and compared those – every single one
24:31
of those stream gauges to every single one of our time series,
24:34
summarized at the watershed scale. So, in doing so, we found that the
24:40
DSWE map time series, summarized at the watershed, when we compare
24:44
those to discharge, we find that the correlations are 0.7 or higher
24:48
for 80% of the watersheds in the Central Valley.
24:51
So that pretty much means that we’re doing a pretty good job.
24:55
If we swap out DSWE for JRC maps, that correlation goes even – increases
25:01
where we get 86% of the watersheds having correlations of 0.7 and greater.
25:06
So strong relationships. If we had poor relationships, that interpolation
25:09
process that I mentioned before, and I will – I’m about to show you
25:13
wouldn’t be worthwhile. We would be doing interpolations of bad data.
25:17
But those strong relationships were a relief because they ultimately allowed
25:23
us to realize that we could do those interpolations with confidence and
25:27
predict – and estimate surface water area where we have data gaps.
25:30
So this is a representation of what a data gap – how the time series
25:34
looked like – looks like for one of the most – for the worst-performing
25:38
[inaudible] in the Central Valley where we had a lot of data gaps
25:41
due to cloud obstruction. And, based on that relationship to a –
25:44
to a discharge record, we can go ahead and interpolate
25:47
those areas in red in the bottom left.
25:50
By having a complete time series like that, which is only 12 years, we can
25:53
go ahead and do a trend analysis. And, in many of the – in many cases,
25:56
in many of the watersheds in the valley, we find that we have 30 years of
26:01
semi-interpolated surface water area estimates. And this is the general trend.
26:05
We see increasing water trends in the Central Valley. And the areas
26:10
with asterisks are indicated as least statistically significant.
26:13
We do seasonal tests, monthly tests, and annual tests for trends.
26:18
All this was summarized and reported in a publication that
26:21
we published last summer.
26:23
And it immediately kicked off two new lines of inquiry.
26:27
One was, how do we identify floods in this really dense time series?
26:31
We don’t just – we don’t just close our eyes and point.
26:35
We have to figure out a systematic way of identifying floods.
26:38
And so we do what we call a z-score analysis.
26:40
So let’s think back to that Friant Dam example where we were looking at
26:44
March snapshots. If we assume that there’s an average
26:48
surface water level at Friant Dam, and then we assume there’s a normal
26:52
distribution around that, where you get a little bit more some years and
26:57
a little bit less in others, then that’s kind of the normal range what
27:00
surface water in March looks like at Friant Dam.
27:02
Now, if you get outliers in the upper limit, that might be an indication
27:06
that something extreme is happening, like a flood.
27:08
If you get outliers in the lower limit, then that might be an indication
27:12
that you have a drought. So those are two ways that, once we
27:16
see a z-score flash an outlier on upper or lower limits, we can go ahead
27:19
and pull those maps at that particular month and go ahead
27:23
and see if we can identify the actual flood event that’s taking place.
27:27
The other effort that the first paper kicked off was a test to see whether
27:32
or not we could transfer and scale that cloud code block
27:36
that we created to move to different geographies.
27:41
So here are the first two places we decided to test that.
27:44
First was Cambodia. And you may ask, why Cambodia?
27:47
We’re a U.S. agency. Don’t worry. We still do the
27:50
United States, but Cambodia was important because it is perhaps
27:54
one of the most flood-prone parts of the world – Southeast Asia,
27:59
low-lying, very similar geography to Vietnam.
28:03
And so, there you get a lot of seasonal variation between
28:07
the dry and wet season in the surface water extents and
28:09
a lot of reported flooding events as well.
28:11
It’s also important because it’s gauge-poor.
28:14
So it doesn’t have the same network we have in the United States.
28:17
And so satellite image and satellite image time series are oftentimes the
28:21
only tool that local governments have to understand surface water dynamics.
28:25
So pretty important to produce maps for a community like this because it
28:29
can be very helpful for making some smart decisions for people who are
28:33
responsible for mitigating emergencies or effects of emergencies.
28:38
The United States, it’s important because we have the ability with that
28:41
stream gauge network to do exactly what we did in California and do
28:44
a full-on semi-interpolation to create a continuous record for most of the
28:48
country and do a surface water – a long-term surface water assessment
28:52
for trends in surface water at the country scale.
28:57
We haven’t done as much work wrapping up that work
29:00
in the United States. More time has been spent
29:03
writing up our results for Cambodia thus far.
29:07
So you may know Cambodia for being the location of Angkor Wat, which
29:11
is this feature shown on the left. But I’m less interested in that feature.
29:15
I’m actually more interested in the water body right in front in the foreground.
29:18
That happens to be Tonlé Sap Lake, which is the largest freshwater lake
29:22
in Southeast Asia. Tonlé Sap can change by
29:26
4 or 5 times over – between the dry and the wet season.
29:29
And so much of the area surrounding the lake there in the middle actually
29:32
is rice cultivation areas. And the communities and the
29:37
people that live in Cambodia are well-adapted to these
29:40
rising floodwaters – or, not floodwaters. Sorry – surface water levels.
29:45
However, when that goes beyond what’s normal and that – like,
29:49
going back to that z-score analysis, if it becomes extreme, then that’s
29:53
when those communities are negatively affected or when
29:56
agricultural production is really negatively affected.
30:00
So this happens to be some examples of what the DSWE maps look like at
30:03
discrete points in time, where you see those different levels of inundation.
30:07
And so, Class 1, which is marked by dark blue, is where we have
30:11
the most confidence in water, and you typically get this gradation
30:14
to wetlands as you move out in the poor water area.
30:19
We went ahead and did our typical validation tests
30:23
in Cambodia, just like we did in the Central Valley.
30:27
And we have good news and bad news.
30:29
Well, the cloud issue is more problematic in Cambodia
30:32
than it is in the Central Valley. And so that was a big –
30:37
a big problem for us. And, in addition to that, we also found
30:40
that we had a lot of missing imagery. And apparently, the Landsat program,
30:44
generally speaking, has data gaps in many of our international collections
30:48
because our international partners didn’t archive imagery the
30:52
same way we do here. So we lost some imagery.
30:55
So, between the missing imagery and the cloud obstruction, over half
30:58
of our months were lost. So we went down from 372 to roughly half that.
31:04
The good news is, the months that remain still have lots of
31:07
good information there. So, while we may be omitting lots
31:10
of floods, there’s still lots of good water information, including
31:12
lots of floods, that we can flag in the months that we have good data.
31:17
In doing what we did in California, where we compared our products
31:21
to the JRC maps, we saw that our correlations
31:23
were 0.78 or greater – 0.78 at the country scale.
31:28
We also did an independent accuracy assessment for 1989, 2015, and 2018.
31:33
And all those tests came out with a result that showed that our maps were
31:38
85% or greater in terms of – or, 85% accurate or greater – pardon me.
31:45
So those maps were solid, and we can go ahead and use those for that
31:49
precipitation comparison if we’re doing comparisons and correlations
31:53
to precipitation data. So this is just a plot of precipitation
31:57
summarized at the country scale monthly compared to our maps.
32:03
And this isn’t a very telling graphic, but you do see that a lot of the peaks
32:07
in precipitation coincide with peaks in our data.
32:10
There’s other instances when that doesn’t occur.
32:13
But we do see some correspondence there, so we do that precipitation is
32:17
having an impact and definitely does – is one of the reasons why surface
32:24
water is fluctuating within Cambodia as a whole.
32:28
So we wrote these data up, and we just got our peer reviews back.
32:32
So this paper is hopefully going to be published very soon.
32:37
But this kicked off our new line of inquiry.
32:39
And this is [chuckles] – so, as you probably heard, I don’t –
32:42
if there is a word of the day, and you screamed every time
32:44
I said “clouds,” we probably would all be hoarse.
32:48
I know I feel hoarse from saying it.
32:50
But we need to figure out a way to see around these clouds.
32:54
And so, of course, we can bring the Sentinel information to bear.
32:57
We could maybe get cloud-free looks in the last six years.
33:00
And that’s a – it’s totally a reasonable path to follow.
33:06
But our thought was, well, we have that MODIS record
33:09
that collects images every day. Why don’t we go ahead and try to
33:13
create these daily time series? Because our thought process is, okay,
33:17
you can wait for Landsat and just hope and cross your fingers that every –
33:21
eventually, at the 16-day mark, when Landsat passes over,
33:25
that you’re going to get a clear view of the landscape.
33:27
Or you can look at 16 MODIS images collected over that same span of time.
33:31
You’re destined to get a clear day, right?
33:34
That’s our hope is that, just by throwing so much data at the problem,
33:37
we’ll get maybe one, two, three clear days that will allow us to
33:41
create monthly composites using MODIS data and maybe
33:44
even create sub-monthly composites.
33:48
We have to do a fair amount of pre- processing of MODIS imagery to kind
33:51
of get it – to create DSWE maps that are comparable to Landsat products.
33:56
And I’ll spare you on some of these heady details, but essentially,
34:00
the satellites pass over the landscape at different times.
34:02
So we have to do some look-angle corrections, brightness corrections
34:07
for different hill shade calculations. We also want to spatially downscale
34:11
a coarse MODIS image to something that’s a little more finer scale where
34:15
we can actually hopefully identify finer-scale features on the landscape.
34:20
We also have to do daily cloud masking within the MODIS product.
34:24
All are very challenging. And fortunately, I have a great team
34:26
of people to help me with this. But we’ve now managed to create
34:30
a MODIS – clean MODIS image time series for which we can create
34:34
these DSWE maps. So now I’m going to show you a few examples.
34:39
These – I did have videos earlier today. It was not working the best, so I’m
34:42
just going to just describe the story. So we’re all California residents.
34:46
We probably can think back to just three years ago, January of 2017.
34:50
Close your eyes. Imagine it. Well, there was a big atmospheric river
34:54
that came through the Bay Area and actually big chunks of California.
34:59
And one of the areas that was affected by that January storm
35:01
was Gilroy, California, where there was a –
35:06
the rains led to floodwaters that actually inundated part of 101.
35:11
This happens to be a news report, if you want to look up the old video, where
35:15
they were just – KPX was just reporting how bad the floodwaters were in Gilroy.
35:21
Well, if we look at the JRC time series, I have a – this is baseline for reference.
35:27
When you look at January, you say, wait, there’s no water in January.
35:29
Well, that’s not the case. You actually don’t have any data in January.
35:33
Cloud obstruction was so great that the JRC folks
35:35
from Europe decided not to create maps altogether.
35:38
And if you look at – in February, you might say, okay, well,
35:41
there are some water features that aren’t there on the baseline image.
35:44
Especially that image in the – around the middle.
35:46
And it’s actually – I should have given you a little background,
35:49
and I didn’t label these. 101 cuts from northwest to southeast.
35:53
Gilroy is smack-dab in the middle. And that blob of blue in the –
35:57
in that February image is the Gilroy wastewater treatment plant.
36:02
So that’s not a flood. That’s just the wastewater treatment plant.
36:06
Now, if we looked at the DSWE time series that we use – that we created
36:09
using the Landsat imagery, we get a little bit of a better understanding
36:14
of what’s happening on the landscape in January and February.
36:18
But if you looked at this data in isolation, you didn’t know anything
36:22
more about – you actually didn’t – so let’s just say you didn’t know
36:25
that the atmospheric river came through the Bay Area in January,
36:28
you might conclude that the flood occurred in February.
36:32
Well, that’s not what we’re trying to do. Because we knew that there was
36:36
a major precipitation event in January of that year, and that’s
36:39
the flood we’re trying to identify. So maybe a little inconclusive.
36:44
Now, if we look at that MODIS time series, where we’re taking
36:47
30 different overpasses to create a maximum water level composite
36:50
for January, we say, well, that’s it. That right there in the middle is
36:55
the big chunk of water that bisects the 101 freeway.
37:01
That’s the event we’re looking for.
37:03
So now that we’ve captured a flood event that we didn’t capture on Landsat
37:06
record, we can go ahead and use that to tie back to that precipitation event.
37:11
Let’s look at another example from far northwest California along
37:15
the Eel River up in the Humboldt Region by Fortuna, California.
37:19
Well, we have a USGS camera at Fernbridge, and bottom left is
37:25
what the January 11th event looked like at Fernbridge.
37:30
If you look at the reference image to the top left from our camera,
37:34
that’s what it normally looks like, where those fields are not inundated.
37:38
And if you have time, this video I happened to find – I don’t know
37:41
who filmed it, so I’m not – I guess I’m not breaking any copyright
37:46
restrictions since I’m not playing it. [laughs]
37:48
It’s a great drone – a bunch of drone footage that shows you
37:51
how extensive this flood was along the Eel River.
37:55
Well, if we look at the JRC maps, once again, no data in January.
37:58
So it can’t be of much use to us. We do see a fork of the Eel River
38:03
pop in in February, but it’s really not enough to tell us the complete
38:08
story of what’s happening. Same goes for the Landsat DSWE
38:12
product, where we have some level of inundation if you look at those
38:15
moderate- and low-confidence water classes, but the fact that
38:20
we don’t have a February scene doesn’t give us any information
38:23
on how water levels may be receding after an event.
38:28
You do see Arcata Bay. Arcata Bay is the feature that does
38:30
pop in inland from the coastline, which is running northeast to southwest.
38:38
But if we look at that MODIS time series once again, we’re seeing that
38:41
level – the Eel River shows really nicely in that January scene.
38:46
So that’s the event that we wanted to identify in January, and MODIS
38:50
is really helping us get there. So a little – we’re kind of
38:53
shooting around. We’re now at the Yolo Bypass
38:56
just west of Sacramento, which is a flood control – so this is
38:59
actually very analogous to Cambodia. Rice cultivating region.
39:03
Flood control site for much of Sacramento.
39:06
So this inundation – you’ve probably all driven over it when it’s been inundated
39:08
and when it hasn’t been inundated. So this is probably not important
39:11
to you. This is all images collected for – not imagery – images collected
39:15
from the Sacramento Bee. Well, January 2017 was a little extreme.
39:20
And so the local communities were actually negatively impacted
39:24
by this event. So here, JRC, once again – it’s not really worth
39:29
spending much time on this one. DSWE produced with Landsat –
39:34
we do see some levels of inundation. And the z-score may flag these as –
39:41
may flag this as a flood. But it’s so cloudy, we don’t
39:45
necessarily know what the full extent of the flooding was.
39:48
Both in January and in February.
39:51
Because, once again, we’re looking at one to two images to make the January
39:55
composite and one to two images to create that February composite.
39:57
So if it was cloudy, then we’re just out of luck.
40:01
Well, the MODIS there – you see how expansive that inundation
40:05
event is in January at Yolo bypass. Especially when you compare it
40:08
to the baseline image. And there, you – that’s the
40:12
event we want to capture. And that’s the one we want to
40:14
tie back to the precipitation that was attributed to it.
40:17
And this last one I threw in there because the Russian River is my
40:23
first personal memory of flooding. Growing up in Redwood City,
40:26
I remember in the ’80s watching the news reports and saying,
40:29
oh, Russian River is going to flood, and not really knowing what that
40:31
meant as a young – as a young child. But knowing that, like, this is different
40:36
than what I normally hear during – watching the news.
40:40
And so Russian River, as we all know as residents of the Bay Area,
40:45
is repeatedly in the news every year because it is so – because the banks
40:49
are oftentimes exceeded, and the local communities are negatively affected.
40:53
So 2017 was no different. And actually, this last winter
40:56
would have been a good one to use as an example because,
40:59
once again, the Russian River and the community surrounding
41:01
the Russian River were unfortunately affected.
41:04
This happens to be – the Russian River is the feature to the left there –
41:08
that kind of linear feature, [inaudible] to the right.
41:12
And so the – so I wish I had labeled this. Once again, I apologize for that,
41:16
but it’s essentially [inaudible] all the way up to Healdsburg.
41:20
Well, if we look at JRC, we don’t really get [laughs] –
41:23
it looks like there’s less water in February than the baseline conditions.
41:27
So that just doesn’t make sense. We know that’s garbage.
41:30
If we go ahead and look at the DSWE Landsat time series,
41:33
we see that event pop in. We were very pleased to see this.
41:38
And I had actually pulled this up before we ever created those MODIS products.
41:41
And I was, like, okay, great. Well, at least the MODIS –
41:43
this DSWE product is doing good. It captured some of these events.
41:47
Unfortunately, we don’t get the full story on the water level recession,
41:49
but we’re still pleased that we can flag that event.
41:53
But the story is more complete when we look at the MODIS record.
41:55
We see that event pop in. We see the water levels
41:58
receded in February. And I just – it really makes me
42:02
satisfied – as I said, rather than – I don’t know if I actually said this
42:05
or I was just thinking it, but rather than counting sheep at night,
42:08
I’ve been counting clouds worrying about the cloud problem at home.
42:11
And so it just really is driving me nuts.
42:14
So just to see the success that the MODIS time series – it has generating
42:19
these winter scenes definitely allows me to sleep a little easier.
42:24
So I’d like to say that, in many – at least in these four examples
42:28
we looked at, the MODIS time series is really great in terms of helping
42:32
fill in that monthly time series and helping us create actually
42:35
a more complete monthly time series for the whole Central Valley.
42:39
So that’s likely to be our next paper is to speak to the full time series
42:44
and hopefully use – do a more robust analysis comparing
42:47
those records to precipitation that’s taken place in California
42:51
over that same 30-year span of time.
42:55
I have, I think, a few – two slides left. So, while we can go ahead and take
43:01
those 30 images – those MODIS images for a given location and compile a
43:05
single monthly composite and fill in that monthly record like I’ve –
43:07
like I’ve said we are now successfully doing, the real strength of MODIS
43:12
may be in that we can actually create a temporal – a temporally
43:15
richer set of maps. We can create 15-day composites or 10-day
43:21
composites or even five-day composites.
43:23
Of course, the five-day composite’s ability to see the – you might have
43:26
cloud problems if you imagine you might have five consecutive days
43:29
of clouds that might obstruct the view of the landscape.
43:32
But, by looking at these sub-monthly scale products, we actually start to say,
43:37
okay, well, at least we can – we don’t know exactly when the peak water
43:40
level was, but we can say that sometime in the middle of the month
43:43
is when we had peak water level at the Russian River site.
43:47
And, by knowing that it’s the middle of the month versus the end of the month,
43:51
we don’t – we don’t have to do a coarse analysis and say, let’s look
43:54
at 30 days of precipitation and compare that to our product.
43:57
We might say, now we only need to look at 15 days.
44:00
Or, now we only need to look at five days of precipitation
44:02
leading up to that event. So there we can create a stronger
44:05
statistical relationship to discrete precipitation events and see how
44:09
discrete events may contribute to discrete flooding events.
44:14
So here are the – here’s one of the conclusions I hope you have
44:18
drawn from this presentation. A, we’re getting closer.
44:21
I mean, we're there, where we now can create a completely –
44:24
complete record of monthly surface water. And, in many cases,
44:27
a sub-monthly record of surface water by bringing MODIS into the picture.
44:33
By having fewer time series gaps, we now have fewer omission of floods.
44:37
So we have a more comprehensive flood record.
44:39
And, while stream gauges can give us indications when floods occur,
44:42
now we actually have the spatial extent of those floods, and we have
44:46
30 years or 40 years, potentially, of that information that wasn’t available
44:52
until these products were created. And, of course, the JRC maps
44:56
were available, but, you know, this is one of the few projects that is
45:00
creating a product that actually gives you the spatial extent of flooding.
45:04
By having a more complete record of floods, we can create a – we have a
45:07
stronger empirical basis with which to establish those causal relationships.
45:11
Meaning we can create a stronger link that’s to precipitation.
45:15
And here’s where we ultimately aim to go by having that strong relationship.
45:18
Let’s just imagine, hypothetically speaking, that we learn from having
45:22
comprehensive records of flooding in Menlo Park –
45:25
this happens to be San Francisquito Creek,
45:27
which borders Palo Alto and Menlo Park.
45:29
This is that 2017 event that I took a photo of.
45:33
And this is a 2019 atmospheric river of a much smaller magnitude.
45:38
But two different types of atmospheric rivers.
45:41
Let’s imagine we have a whole record of flooding for Menlo Park.
45:46
And we know from that record that 3 inches of rain leads to considerable
45:50
flooding within the community. Whereas, 1 inch of rain may only
45:53
lead to – let’s just say it doesn’t lead to any flooding whatsoever.
45:57
By having those – knowing that, we can say, okay, well, somewhere
46:01
on that continuum from 1 inches to 3 inches, we’re going to get – we’re
46:04
going to get flooding in the community. Now, at the 2-inch mark, we may say,
46:07
okay, well, these parts of the community are going to be affected.
46:09
Wouldn’t that be valuable for local – for local government to say, okay,
46:14
well, now we can put all the sandbags here rather than just kind of spreading
46:17
them all thin around the community. This actually can help identify
46:21
those communities affected and actually help prioritize mitigation.
46:26
So that’s eventually where we want to get with this work.
46:28
And with that, I should have acknowledged these people
46:30
at the beginning of the talk.
46:33
Because they are so critical in the success of this project.
46:35
But I’ll acknowledge them now. Jessica Walker, Roy Petrakis,
46:38
and Eric Waller, who have all been integral parts of this project,
46:41
and I really appreciate – I really appreciate having them on my team.
46:45
And with that, I think I have some time for questions. Thank you.
46:48
[Applause]
46:55
- Please use the microphone. [laughs]
46:57
- [inaudible]
47:00
[Silence]
47:04
- Hi. Thanks for your talk.
47:06
Maybe you touched on this, and I missed it, but what –
47:10
you know, when you showed some of the images we just showed –
47:13
a photograph, you know, of visible light.
47:15
And I know you have multi-spectral capabilities as well, or at least
47:18
you alluded that you did. - Mm-hmm.
47:20
- So what does the pixel classification – you know, you get a pixel.
47:24
You get pixels back on your image. You figure out, is this vegetation,
47:27
is this water, is this cows, or whatever. How is that process done?
47:32
- Oh. That’s – I’m happy you asked. I thought it was going to be a
47:37
put-people-to-sleep slide, but I … [laughs]
47:41
So the dynamic – well, let me say, the JRC process, it’s somewhat
47:47
unclear how they go about their mapping of surface water.
47:51
So I can’t really speak to that. I can speak to the Dynamic
47:54
Surface Water Extent model. So what they – what that model
47:57
does is actually does – creates a bunch of different water indices
48:02
using all different types of spectral bands.
48:05
So sometimes – some of those indices rely on visible parts of
48:10
the electromagnetic – or, the visible parts of the spectrum.
48:14
And some use infrared parts of the spectrum.
48:16
But essentially, they create a variety of water indices.
48:19
And hill shade – and a hill shade component as well.
48:22
And then it creates a composite score. So it says, like, okay, this pixel met
48:26
this test, it met this test, it met this test, and it’s within –
48:29
it’s under a slope of 30 degrees – because the model assumes
48:33
that if you’re on a steep slope, you can’t have standing water.
48:36
And so, if it meets all those tests, then it gets flagged as
48:39
a high-confidence water pixel. If it meets four of those tests,
48:42
then it may be flagged as a moderate-level water pixel.
48:45
So I’d be happy to share with you that documentation if you’re interested.
48:48
But I think it uses four or five different water indices to ultimately make that –
48:54
it’s a decision tree where it makes a decision
48:56
based on how the composite score bears out.
49:00
Question in the back?
49:03
[Silence]
49:06
- Having been the area where you talk about – Cambodia –
49:10
under different circumstances, of course, why Cambodia,
49:15
of all the countries in Southeast Asia, did you choose to compare to the
49:21
flooding levels or the rainwater levels or the surface water levels here?
49:25
What was so unique about Cambodia of all the countries in Southeast Asia that
49:30
you chose it to do a comparison study? - That’s a good question.
49:33
So I guess I could have chosen any country in Southeast Asia.
49:41
I don’t have a real good answer for you. I mean, I think I tinkered
49:44
a little bit with Vietnam. I tinkered a little bit with Thailand.
49:47
And I just – for some reason, Cambodia seemed, like,
49:50
a little larger in size, so I felt like it was a bigger piece to bite off.
49:55
And so I – so I don’t have a real good rationale for why I chose
49:59
Cambodia over any other country there. Because I know all those countries
50:03
have limited gauge networks. So they all kind of meet that same
50:05
criteria of being flood-prone and gauge-poor, more or less.
50:10
Yeah. I wish I had a better answer for you.
50:13
- [inaudible]
50:15
- Well, I mean, as with most of our work, you know, we start creating –
50:22
with this – with this cloud-based code – pardon me – cloud-based code,
50:27
I could go ahead and create a couple examples of surface water
50:30
and surface water change. And I saw how much Tonlé Sap Lake
50:35
changes in its extent. So I was, like, well, this is kind of fascinating.
50:38
I didn’t even know what Tonlé Sap was before I started this work,
50:42
to be quite honest. And so I said, okay, well,
50:45
let’s go ahead and just crank these data out and see if they actually
50:48
produce a nice, long time series. And came back to my computer the
50:53
next day, and I started tinkering with it. And I said, well, I think we’re
50:56
onto something. And I never second-guessed it
50:58
and went to try another geography, so, yeah, I think that’s all
51:02
that really led me down that path.
51:09
Any other questions?
51:12
- [inaudible] - Oh, no. Of course.
51:14
- [inaudible]
51:17
- I’m sorry. Would you – I know that they’re recording this,
51:19
and they probably want you on a mic.
51:22
- I saw on your screen that there was few imagery of Cambodia and that
51:29
the library pack, okay – the library pack was very low.
51:35
First, you don’t – you're gauge-poor, I think you said – gauge-poor.
51:40
- Mm-hmm. - Imagery-poor. Measurement-poor.
51:45
So I was – I was saying, well, I’ve been to Vietnam. That’s – why not Vietnam?
51:50
Because we’ve been there. Okay? And the cloud cover you
51:53
were talking about – I said, I know that. Okay. So …
52:01
I’m just wondering, why was a science-poor area like Cambodia
52:07
used to a science-rich area – like, say, the Bay Area.
52:11
When you were talking about Gilroy, where I’m from.
52:14
Or you talk about Yolo. You talk about Russian River.
52:18
Where I’m from also. So I was just trying to follow you
52:23
and say, okay, there’s Cambodia. And there’s us.
52:27
What’s that about? Sorry. - Oh, no, no, no.
52:32
So, like I said, picking of Cambodia was driven initially by two criteria.
52:40
We wanted to go to an area that had more floods than the Central Valley.
52:43
We wanted to – and we wanted an area that was gauge-poor because we knew
52:46
how important this product could be in those sorts of communities.
52:50
So, as we touched on a moment ago, that could have been Vietnam.
52:53
That could have been lots of different – Malaysia.
52:56
We could have – we could have chosen a lot of areas.
52:58
But really, this is a pilot project to showcase our ability to
53:02
roll this out anywhere. So, now that we’ve done this, and
53:06
we’ve actually published that code,
53:09
you, this gentleman right here – anyone can move that code over and
53:12
produce Vietnam at a click of a button. And so you can go on my data release,
53:17
if I made it available on here – I’m sorry for not doing so – and we can all create
53:20
these data – they’re a little hard to look at after you’ve created them,
53:24
but we can essentially all create – we can create a global data set
53:27
using the same code block. So whether or not we tested this in
53:30
Germany or Cambodia, it ultimately doesn’t really matter because now that
53:35
we’ve proven that it works in a different geography, now we can roll
53:39
this out anywhere you would like to. But I just thought it was
53:42
of high value for Cambodia. And I wasn’t making a case
53:45
that it was more valuable to Cambodia versus another government.
53:48
I just settled on Cambodia.
53:53
[Silence]
53:59
- I think I saw on your slide showing the Central Valley surface water,
54:06
that there was an increase over the last decade or so.
54:09
- Mm-hmm. - Some of that, I assume,
54:10
is precipitation-related, but there’s also a lot of
54:12
groundwater pumping going on to support agriculture there.
54:16
Are you able to distinguish in any way what influence the groundwater is
54:21
having in – that’s kind of leading to the question in the future.
54:24
Are you going to be using GRACE data more in your research
54:27
to examine what’s truly happening for the water people?
54:31
- That’s a wonderful question. And I imagine many of you probably
54:33
saw the 60 Minutes piece from a few years ago that talked about
54:37
pumping in the valley. And we obviously –
54:39
it’s also in the news quite often around here.
54:44
This is purely surface water. So I haven’t – yeah, I haven’t gone in
54:47
to try to tell that story about subsurface dynamics.
54:51
Because I largely stay in the realm of what I can see with a satellite.
54:55
There has been lots of work by other USGS scientists in that –
54:59
scientists in that space. But I haven’t done any work in that space.
55:05
And right now, like I mentioned, that work trying to complete the
55:08
time series was our – was our focal point.
55:13
And so that’s been where we’ve been spending our newer more recent efforts
55:17
trying to – trying to work on that and improve what we were previously –
55:22
what we were previously producing in the Central Valley.
55:23
So unfortunately, I haven’t made any headway in that [inaudible].
55:29
[Silence]
55:33
All right. - [inaudible]
55:35
[Applause]
55:42
- Thanks, everyone.
55:43
- Thank you.
55:47
[Silence]