Observing Global Croplands from Space | Wisconsin Public Television

Observing Global Croplands from Space

Observing Global Croplands from Space

Record date: Feb 21, 2018

Mutlu Özdoğan, Associate Professor of Forest and Wildlife Ecology at UW-Madison, discusses using satellites to photograph and monitor crops from space. The satellite photos offer information on crop-management practices and a look at global changes.

University Place Campus: 

University Place Lecture Series: 

University Place Subjects: 

Episode Transcript

- Welcome everyone to

Wednesday Night at the Lab.

I'm Tom Zinnen.

I work here at the UW-Madison Biotechnology Center.

I also work for UW-Extension

Cooperative Extension

and on behalf of those folks

and our other co-organizers,

Wisconsin Public Television,

Wisconsin Public Radio,

Wisconsin Alumni Association,

and the UW-Madison Science

Alliance, thanks again

for coming to

Wednesday Night at the Lab.

We do this every Wednesday

night, 50 times a year.

And today, is the

anniversary of our

12th year of doing

Wednesday Night at the Lab.

Next week, we start

our 13th year,

so thanks for coming out

every Wednesday night.

I appreciate it.

Tonight, I'm delighted to

be able to introduce to

you Mutlu Ozdogan.

He's with the Forest and

Wildlife Ecology Department.

He was born in

Ka-tyr-a-da, Turkey.

And I am fluent in Turkish. [audience chuckles]

And that's also where

he went to High School.

He went to Istanbul

University and he corrected

me as I was writing down

the word, In-stan-bul.

He said, "There's no

N in In-stan-bul".

Well, there's not two Ns.

And for my entire life,

I've been pronouncing it

"In-stan-bul".

This is what we

call, adult learning.

[audience chuckles]

That's where he got

his undergraduate

degree in Earth Science.

Then he went to North

Carolina State University

to get his Master's degree.

Got his PhD. in geography

at Boston University.

Then he post-doc'd

at this little place

you may have heard

of called, NASA.

Ten years ago, 2007, he

came here to UW-Madison.

Tonight he gets to talk

to us about mapping

the world's croplands with

this great view from up above.

Not just a bird's eye view,

but a satellite's view.

Please join me in

welcoming, Mutlu Ozdogan,

to Wednesday Night at the Lab.

[applauding]

- Thank you, Tom.

- Thank you very much.

- Is this hot I think?

Well, thank you so much

for coming, I really

enjoy these events.

I think it's a great service

that the University provides.

This is my second

Wednesday Night at the Lab event,

and back then, you

know, in the first one,

I talked about

looking at forests.

Now, I'm gonna talk about

looking at croplands.

Before I start my presentation,

I wanna ask this question of

why agriculture is important.

And I'm sure people have

different views about that,

but these are some of

my views over here.

Obviously, agriculture

provides food and nutrition.

You know that's something

that's the stuff that we eat.

But many people

around the world,

agriculture is also

a livelihood, right?

There's a picture of

someone trying to extract

some sort of yield

from the field so that

they can sell that and

then make money off of it.

Increasingly, the agricultural

areas are used for

fuel for growing bio-fuel

products and there's

a lot of research going on,

on this campus.

And I would also argue

that agriculture is sort of

this tradition, which

is associated food,

and obviously economy.

You know, we're in the Midwest.

Around here agriculture

plays a big role

on the economy.

And it turns out that

all of those things are,

in my view, interlinked.

So, if I put the cropland area's

defined as sort of

where crops are grown,

a variety of crops, all

of these things that

I just talked about

are interlinked. Right?

So the food is related

to cropland industry,

nutrition, climate.

How sort of cropland

areas regulates the

climate that is sort of

happening above that.

And then water obviously

population, economy.

So there's a lot

of sort of linkages

that agriculture

brings together.

Not only that, but as

our population grows,

and you know, we

reach somewhere around

nine or 10 billion

people, we do have to

grow more food, but

believe it or not,

Earth is running out of space

in terms of growing food.

We're actually busting

at seams in terms of

the best agricultural land,

'cause we already found those

agricultural land

around the globe

and we're cultivating them.

And if you wanna grow more food,

we either have to increase the

sort of per unit area yield,

which we're working

on, or we need to

expand agricultural areas.

But the only places to

expand are places like

the Amazon forest,

the South American

pristine grasslands,

and so on and so forth.

And not only that,

but agriculture also

has this sort of

interesting balancing act.

This is a slide

that I stole from

the World Resources Institute.

So, not only do we have to

sort of increase the food,

and also help people

bring livelihood

out of agriculture,

but also we need to

do that without wreaking

havoc to this planet.

Agriculture has sort of huge

environmental footprint.

Whether that's removing the

existing sort of natural areas

for to make pay for

agriculture, or to intensive

agriculture that

uses a lot of inputs,

like pesticides and nutrient

inputs that we see around here.

So, I stole these slides

from a colleague of mine,

Jon Foley, who was

here in the past,

and he talks about

how croplands are

prominent across the planet

and covering roughly,

about 40% of the planet.

And not only that, but

to grow that much food,

we also need to

use a lotta water.

So I'm trying to get at this

theme of agriculture is great.

It provides all those services, but it also comes at a big cost.

So the nutrient cost

and the water cost,

and I'm sure

everybody knows this

great story of the Aral Sea, as you can see in this satellite,

in the last 30 years

how withdrawal of water

from the Aral Sea for

irrigating agricultural areas

has led to the

diminishing of that area.

So, agriculture is

a big user of water

and globally about 80%

of global fresh water

resources is used

for agriculture.

So the joke that I have

with my students often,

is that those of them who take

long showers, that's okay.

You're not the culprit,

agriculture is,

with respect to water usage.

Now, I think I've tried to

sort of convince you that

agriculture's

important and it also

comes with some

environmental footprint.

Now, even though we know

that and that's a fact,

we have very little

knowledge of agriculture

at certain parts of the globe.

Now if you look at the

Midwest or the US or Canada,

and perhaps Europe or

Australia, we have very good

knowledge of where

agriculture is

and what kind of crops are

being grown and their

average yield value,

so that we can sit down

and do some sort of

interesting analysis.

But there are parts of the

world where agricultural

information is sorely lacking,

and I would count Africa,

India, and parts of

southeast Asia in there.

So our knowledge of

global croplands,

I would argue, is regional.

Not globally fully complete.

The second argument

I would make is that

many of the agriculture,

global agriculture

data collection agencies,

like the United Nations

and food and agricultural

organization depend on

individual countries to report

their agricultural areas.

And those reports are

often what I would call

subjective reports,

because they don't

necessarily depend on

any objective assessment.

And the planet is

changing really fast,

so we're cultivating

more cropland.

We're changing the types of

crops that we're growing in an

existing cropland area, so

changes happen really fast.

And I would also argue that

more and more environmental

impact analysis associated

with agriculture require maps

or data sets that are spatially

explicit, and accurate,

and up-to-date that we

don't have everywhere.

And finally, these data

sets, if we can get at

these data sets that could

help us shape the future.

So with that, what

I'm gonna do is,

I'm gonna pose the second

question of what is

the role of remote sensing?

Now for those of you who are not

familiar with remote sensing,

the text book definition

of remote sensing is

acquisition of information

about an object without

physical contacts.

Just the fact that you guys

are watching this presentation

and forming some

opinions is actually a

form of remote sensing.

So with respect to

satellite remote sensing,

what we do is we

build these equipment,

they're basically

souped up cameras,

and we put them in

satellite platforms

and we launch them in space.

And as they circle

around the planet,

they take a variety

of pictures in

variety of different colors.

And those colors include

all the visible colors

that we all can see, but also

colors like the infrared,

the short-wave

infrared, long-wave

infrared and thermal-infrared.

And it turns out that

many vegetation studies

that we conduct are

very sensitive to the

colors that are built

into those cameras.

Now there are lots of additional

benefits of using satellite

observations in the form

of remote sensing that is

large area coverage,

objective observations,

because we're not

relying on a person's

opinion but rather

the opinion of light.

A physical quantity.

There are, as I said,

spectral capabilities

that we can see beyond

the visible color

the human eye can see.

And there is often

wall-to-wall coverage.

There's data anywhere

within the planet

even in the inaccessible places,

going back 40 years, so we

can do historical analysis.

And these data sets are

very accurately located

in terms of their

geospatial location.

So that allows us to ability to

acquire a lot of information

in a very short time.

Now even though remote

sensing has these

skills and has been

around for a long time.

And these are some of the

different agricultural examples

that remote sensing can give

you from different parts

of the world, from the

Midwest US to China,

to India at different scales.

Remote sensing has not been very

operationally used for

agricultural assessment.

And there are many

reasons for that.

Even though one of the very

first Earth observation

satellites were launched

because of agriculture,

people immediately realized that

there were a lot of

challenges as well.

So some of those

challenges are mismatch

between the needs

and the deliverables.

What I mean by that is if

you go and talk with farmers,

they wanna see their

own field everyday

with a certain wavelength.

And it turns out

that one of the early

satellites that were

built, simply were not

capable of delivering

that kind of information.

Now, both the public and

private organizations are

addressing that issue, but

that has been a challenge.

And then satellites however,

have been designed for

this multi-purpose design.

Engineers haven't

built these satellites

just to look at agriculture.

They built the same satellites

to look at forestry,

to water, to urban

areas, and et cetera.

So there isn't the ag-specific

remote sensing tools

or assets that are out there.

And then, we actually

don't have the

ability to look

at every variable.

So for example, if you ask

me to tell you how much

fertilizer a farmer

puts on their field,

using satellite data, that's

pretty much impossible.

Maybe we'll get there

in the near future,

but we're not there right now.

And there has been sort

of traditional cost and

processing associated with this,

especially with large

volumes of data.

So, remote sensing has

been around for a while,

but there has been

some challenges.

But, in the last

five to 10 years,

there also has been a lot of

opportunities coming our way.

And I'm happy to

report that because

that's sort of job

security for me.

[audience chuckles]

So, and some of these

developments are related to

computation and some

of it is related to

changes in policies of

how data are delivered.

So for example, when I

started using remote sensing,

15, 17 years ago, you

could only purchase a

single tile of image that

over, let's say Madison,

and maybe at one-time period,

and you would have to

pay 2,000 or 3,000

dollars for that.

And that's the

only data you got.

So the data volumes and

prices were somewhat,

they would inhibit the analysis.

Nowadays, we can look

at the entire planet

and we can look at the,

download the entire archive

of last 40 years of data

and then process that.

But that requires

additional processing.

So some of the opportunities

that I've been seeing

and that I've been

employing in my own lab

are the ability to process

large volumes of data.

The second part is,

there is this very active

and growing open

source community.

There has been a

lot of smart people,

and I'm not one of them in

terms of computer sciences,

have been developing these

really interesting tools

that are directly applicable

to satellite data.

And then, there's also a lot

of active community out there

that is providing science

to open source community

and we're doing some of

that at this university.

Let me skip that.

So these are some of the data

tools that I'm gonna show you

results of in a few

minutes, but these are

the kind of tools

that we've been using.

So what I wanna do is the

title of this research,

this talk is "Observing

Crops From Space,"

so what I wanna do is,

with that introduction,

give you some examples of how

we observe crops from space.

And I'm gonna do that in sort

of three different contexts.

The first one is, we're gonna

look at the entire globe,

and figure out where crops

are, and we're gonna try to

do that at various high spatial

resolution, if you will.

The second one is,

trying to map crop types,

specific crop types, corn, soy

bean, wheat, in large areas.

And I'll show you

examples of that.

And the third one which

is more dear to my heart,

is trying to get at the crop

yields at the field scale.

So, if you look at Dane

County for example,

and let's assume that

there are 1,000 corn fields

in Dane County, wouldn't

it be nice to look at

and estimate yields of

every one of those fields

and then use that information

in a variety of ways.

So let me start

with the first one.

So, about four years ago, we

had a project funded by NASA.

And NASA has a program

called Measures,

and Measures, you know

NASA loves acronyms

and Measure is also an acronym.

And the idea of a Measures

program is that they wanted to

fund researchers that would

utilize satellite data,

mostly NASA data,

and other forms of

data to provide products,

or produce products,

map products, for societal good.

So we submitted a proposal,

and we got a proposal accepted.

And the idea of that proposal

was to map global croplands

at different scales and regions

at 30-meter spatial resolution.

And those of you

who are not metric,

30 meters would

be about 30 yards,

so each pixel would see a little

less than a tenth of an acre.

So imagine dividing the

entire Earth into less than

tenth of an acre plots,

and then trying to

map whether there

is an agriculture or

not agriculture

in that particular

plot for the entire

global landmass.

We called that product GFSAD30,

so Global Food Security Support

Analysis Data at 30 meters.

A mouth full.

And we've been leveraging

this idea of cloud computing

because of just sheer

volume of data,

and those data sets

are being delivered

right now as we speak.

The whole goal of that project

was to produce

four map products.

The very first product

was to simply tell people

whether a particular

piece of land was

cultivated or not cultivated.

So a binary mask, if you will.

Now once we derived

that particular map,

then the next idea

was to, sort of, driving this

map of cropping intensity.

Now, cropping intensity,

now in this sense,

simply means the number

of cycles, growing cycles

within a calendar year, or

within a growing season.

Obviously around here,

our growing cycles

are only one because we

have a short growing season.

We can only fit in

one crop barely,

if you will, around here.

But there are places

around the world where you

can grow two, three,

sometimes four crops.

So this second product

is looking at the

number of vegetation cycles,

number of growth cycles

of agriculture in that

particular parcel of land.

The third one, which is probably

a more difficult product

is a crop type map, in

which we would tell you

what is the dominant

crop type in that

particular parcel of land

that I've talked about?

And I use the word, dominant.

That's because every year,

crops parcels change.

So around here, if I

was to create a label

for a parcel of land,

that label would look

something like

corn soy dominated.

And in North Dakota, it

might be wheat soy dominated.

And in parts of

India, it might be

rice dominated agriculture,

because of their

changing nature of agriculture.

And final product that

we were interested in

driving that we're

delivering right now

is this idea of irrigated

versus non-irrigated.

So it's taking a look at

the same parcel of land

and tell you whether

that piece of land

receives irrigation

or not, and there are

some abilities in remote

sense satellite data

that would allow us to

detect whether that piece of

land is being artificially

irrigated or not.

And the reason that's

really important is because,

as I said, globally 80%

of fresh water resources

are used for irrigation,

and only about

18% of global agricultural

area is irrigated.

But the yields in those

locations are really high.

So for this presentation,

I'm only going to

show you the examples of the

very first product up there.

And then some

examples of the third

product in a later part

of the presentation.

Now the way that we

approached this problem is

we had a large team,

and each of those

team members had experience,

geographic experience,

in different parts of the world.

So a team took North America.

Another team looked

at South America.

My team looked at Europe.

You know, I'm from Turkey,

so I've been in that

area quite often, so

we know that area well.

We took Turkey, Middle

East, and Central Asia.

And then other groups

have looked at different

parts of the world,

and what we did is

we used slightly different

computer algorithms

to drive the maps that

I was telling you about,

and simply stitched them

together to make a global map.

And then delivered that

to public resources

that you and anybody

else can download.

The way that we

approached this problem

is by taking satellite data.

So this is a mosaic

of a particular

satellite data over

Africa, shown in what

we call a false color composite.

And false color composite

simply means any place

we see a red color would

be actively vegetated.

That's why it's

called false color.

And this is actually

a term that is

left over from

the World War Two.

Then we could take

those indices data

and turn them into

vegetation indices.

So, in this case,

the more green it is,

the more actual

vigorous green it is.

But not only that,

but we could also

take additional variables,

like topography,

slope, and et cetera, and

then bring all of these

into a computer machine

learning algorithm

in which you provide examples

of what cropland looks like,

and simply computer finds

very similar examples for you.

But instead of doing it for

one pixel or one location,

it does it for billions

of billions of locations

in those little parcels

that I talked about.

Now I don't know

if you can see it,

but if you look at the

middle part of that

very left picture,

you can actually

being to see little

square boxes.

I don't know if

you can see that.

And the reason you

are able to see those

is because those are

the individual tiles of

satellite image that

is being collected

with the satellites

that we're using.

So in this case, we have

to take many of them,

in this case, literally

tens of thousands of them,

and then clean them for

clouds and other issues,

and stitch them together to

make these mosaics so that

they can be used in an image

classification algorithm.

Now the type of classification

algorithms we use

were what's called a supervised

classification algorithm

in which an analyst,

like myself,

would have to provide

examples of what a

cropland looks like to

a computer environment.

And those are not that

easy, especially when you

consider a global agriculture,

because global agriculture

is highly variable.

And how do you get the sample?

So we reached out to global

organizations like the FAO,

the United Nations,

and other organizations

that helped us collect

literally thousands of

on the ground examples of

what agriculture looks like,

so that we can take those

examples, and then feed them

into a smart computer algorithm

so that those algorithms

could recognize what

agriculture looks like.

And obviously, that

is really different

across different

parts of the world.

In the Midwest, agriculture

looks very different

than what it looks like in

China or in India for example.

So that was a great effort.

I would say we spent half

of our time just doing that,

instead of processing

satellite data.

So these are some examples.

This is an example from

Australia just within

the continent of Australia

or country of Australia,

there is a lot of variation.

Everything from rain

fed single crop wheat to

rain fed single crop lupine

to canola and et cetera.

And this is within

a single landscape.

And if you wanna map

agriculture in this landscape,

you would have to need to

know all of these different

sort of character or

characterization of agriculture.

Similar things in Africa?

So Africa was even a

more difficult case,

because in Africa,

there isn't a whole lot

of commercial agriculture.

There are parts of Africa that

have commercial agriculture,

like South Africa,

Egypt and northern part,

and some in Kenya.

But there are other parts

of Africa where agriculture

is just sort of a very

transient in livelihood.

And then how do you

collect sample so our

team members from the UN and FAO

has been instrumental in getting

to the field and collecting

those samples for us.

Now once we took those samples

and then ran them through

a computer algorithm, along

with our satellite data,

this is the kind of

map that we produced.

So you can go to that

website and you can

download and you

can look at these.

I think it is also mobile able.

But all the green,

this bright green areas

show the places where

global agriculture is.

So what I want to do

is, so this is the first

product that I was

talking about earlier.

So what I wanna do is, I wanna

take a few places around this

global map and show you

examples of detail product.

So here's an example

from Bulgaria.

So it's right there,

for those of you

who don't know where

Bulgaria is, right there.

So, here is what a

global Google Earth

image background

would look like.

Here are some mountains

with dark forests,

and then this is where

the agriculture is.

And our computer algorithm

capture where agricultural is

and not capturing

anything that is

forest or urban or

roads and et cetera.

So that's one example

from Bulgaria,

sort of a temperate

location, if you will.

Here's an example from

Kingdom of Saudi Arabia.

So Saudi Arabia

15, 20 years ago,

decided that because of

the ground water resources

that they have that

they were gonna

grow a lot of wheat and

then export that wheat.

What a crazy idea.

So this is one of their

big project areas.

So all of those are large

scale irrigation projects,

and our product captures

accurately those.

And I would say

that those are easy

examples in Kingdom

of Saudi Arabia.

Then here is an example

in Vietnam, where here

is the river.

This is in the Mekong Delta

area, here is an urban area.

There are some roads,

and there is a lot of

rice cultivation

between those areas.

So you can do this yourself.

You can go it, to that

particular website.

You can look at it

on a mobile phone or

iPad or a computer,

but those maps exist.

And we are very proud

that we were able to

produce products like

these that are of

service to a variety

of organizations.

Now, when I teach

remote sensing,

I tell my students to

promise me two things.

And those are, never take

a map from someone without

quantitatively

assessing its accuracy,

'cause anybody can

make a map. Right?

Making an accurate

map is important.

And the second promise

is never deliver a map

without first assessing

its accuracy and

delivering that product

with that accuracy number.

So we spent a whole lotta

time validating our map,

and validation simply occurs

by going to different places or

extracting samples

that are independent of

the training samples, and

simply trying to figure out

how much of those

samples that are on the

landscape have been accurately

mapped by the algorithm.

So here is a sample data set.

So we developed some

mobile tools to

look at very high

resolution imagery on an

iPhone or on a mobile device,

and then simply put a

dot in the middle of it

to figure out whether this

is agriculture or not.

And we've done that over

thousands of locations,

and in some locations we had

access to other data sets.

That's why there's this

important break over here.

And then we compared the maps

that we made against those,

and I'm happy to report

that on a global average,

that particular map

is about 90% accurate.

Now that's a global

average number.

There are places where

the map is 99% accurate,

and there are places where

the map is only 60% accurate.

That's just the nature of

agriculture and how it exists in

that location and how remote

the sense data may or may not

be able to pick it up in

that particular location.

I can tell you that in the

Midwestern United States

around here, mapping

agriculture is not that easy.

And that's because

we have large fields.

So agriculture is really

established and you can

really see in the

remote sensing signal

what is agriculture

and what is not.

But if you look at

other places like China,

or India, parts of

India where we didn't

a place as big as this room,

there's probably five different

land cover types,

including several trees,

animals, humans, a

couple of houses.

Agriculture becomes a really

complex variable to determine.

In those areas, map

fails to deliver

at that sort of 90% accuracy.

So these are just

examples of those apps

that I've told you about.

Another way of doing

validation is to

compare our product

to other products.

So in the beginning of

this talk, I mentioned that

even though there are

existing products out there,

they're not always

accurate, or they're not

always capturing the

type of agriculture

that we're talking about.

So this is a map that simply

shows three different products.

So here is a background image.

This is in southeast

part of Egyptian delta.

So here's agriculture,

here's non-agriculture,

and here's some

agriculture over here.

This is a global

product that has

been produced 10 years ago

at 10-kilometer resolution.

So that's a pretty coarse

resolution data set.

And here is a map that has been

produced at

250-meter resolution.

And here is a map that has been produced at 30-meter resolution.

So as you can see, even

though they're all showing

agriculture because

of the sort of

spatial resolution issues,

you don't necessarily get

the right area, or

the right location.

Now the reason I'm

showing this, is because

many of the

global aid agencies,

whether it's USAID

or World Bank or FAO,

has been traditionally

relying on maps that are

made from those, because

those were the only available

products at scale that those

organizations could use.

And they were extremely

excited to see the type of

products that we're producing

at extremely fine detail.

The second topic I wanna

cover is refining the maps

that I described on a

more categorical level.

So, in addition to

mapping a parcel,

as being either agriculture

or not agriculture,

the next level of information

is what is being grown in

that particular parcel of land.

Now that we know that it's

agriculture, but what is

really cultivated with

traditionally, or predominately?

So I wanna show you

examples of that,

and then that's a slightly

different problem.

I would say it's

a bigger problem.

And the reason for

that is two-fold.

One is as the information that

we require from satellites

becomes more refined, like

going from agriculture

versus not agriculture, to

what type of agriculture it is,

the signal that exists in

satellite data declines.

That's because if I was to

show you corn and soybean,

in a satellite image in a

particular time of the year,

you couldn't really distinguish

them from each other.

That's because at that particular time instance,

they are both green,

and we don't necessarily

see the leaves, as I

showed you some of the

spatial resolution issues, you

can't really separate them.

So we need to have these

additional variables

in our possession to be able

to separate those crops.

And it turns out that to be

able to separate individual

crops from each other,

the temporal information

becomes the most important

piece of information.

But that means that when

you process the data,

you have to access this

time series of data,

starting from the

beginning of planting,

all the way to harvest,

for all the crops

that you're interested

in a particular land.

So one example around

here, for example,

is separating corn

from soybean.

As you know, one

of them is planted

a couple of weeks earlier,

and it grows slightly faster,

and it peaks at a vegetation

that it's particular time.

One is earlier

than the other one,

and the senescence

a bit earlier.

So if you have a

time series data that

captures that small difference,

couple of week difference,

you would obviously

separate corn from soybean.

But what happens if

a farmer decides to

plant his or her

crop two weeks later?

Then in that case, they become

somewhat indistinguishable,

and those are the kind of

issues that you would run into.

So, this is a different

image processing

or satellite processing problem.

And the idea over

here is to convert

these satellite observations,

here is an example

of what a true color

composite, where you can

see individual fields,

into a map like this

that shows where the

individual crops are,

and then simply separate

them from each other.

And as I indicated,

the way that we do that

is to assemble this

time series of data.

So, we have observations

that are roughly

observed about a week apart.

So anywhere on the

planet, we can get an

observation on a week apart.

Now when I say week

apart, we get a

potential for observation today.

And then another

potential for observation

about seven or eight

days from today.

Now that doesn't mean the

observation will be made.

That's because Earth

is a cloudy place.

And especially around

here in the summertime,

you get a lot of clouds.

So there might be

times in Madison,

where you don't see the

ground at least with the

type of satellite data

that we use, for a month.

So that means you have this

entire month of data gap

in your observation,

and what if that's the

time of separation of

corn from soybeans?

So what we always try to do,

is we're always trying to

find ways to fill these

gaps, either by using

additional data sets,

which may have different

characteristics themselves,

or try to fill them with

additional gap filling

methods that are

more rooted in mathematics

and computer sciences.

But at the end of the

day, our goal is to

build this time

series and apply that.

Now a few minutes ago, I

said mapping crop types,

it has two problems.

One is assembling

the time series.

The other problem

is how do you get

training and validation data?

Now I could give all

of you a mobile app

and I would be highly

certain that all of you

can tell me whether

you're looking at an

agriculture land or

not agriculture land.

But it becomes much more

difficult to tell me if it's

corn or soybean, just by

looking at the satellite data.

So even though we had 100,000

samples for mapping crop,

non-cropland, we only had

1,000 samples for knowing

where crop types are, in

terms of a global sample.

So there's this multiple fold

decrease in both validation

and training data that we

could fit into our algorithm.

So, I'm gonna show you some

examples from Dane County

because we all know

where Dane County is.

So here is a product

that we didn't generate,

and I'm showing this

just as the example

of what's possible out there.

This is a product that

has been generated

by the US Department

of Agriculture.

And here is a crop type map

for Madison and Dane County,

and this is a crop

type map for year 2015.

And here is year

2016, and obviously,

if you go back and forth,

the last years corn areas

became soy and last year's

soy areas became corn.

That's not always

the case, but that's

the most of the

case around here.

So the reason you're able to

do things like that in a place

like Dane County is because

there are known patterns

that we can extract data

from and then map them.

Now try doing that in Asia,

or in India, or China

where there is no

specific rotation

pattern, or there's very

little ground

related information.

Anyway, that was 2015, 2016

and here's your 2017 data that

just came out.

Now, for certain places

around the world,

we had really good data,

and because some of

our partners were

actually from India,

and they were able to

provide the ground data

so that we can feed

that ground data into

our computer algorithms

and map crop types.

And not only crop types,

but also whether they're

irrigated or double crop

or single crop, crop type.

So I'm just gonna read

some of these for you.

So here is an irrigated ground

water irrigated rice, maze,

and chick pea color, wherever

that color is right here,

for example, that parcel of land has that kind of attribute.

And obviously, in order to

determine that fine scale

or fine categorical detail,

you would have to know

a whole lot of information

about the ground and without

our Indian partners, this

would've been impossible.

So the message I'm

trying to give over here

is that we are able to

make maps like these,

but those maps will vary

highly in their quality

depending on how much

ground data that we would

have access to, and how

much satellite signal,

or how easy it is in

the satellite signal

to define that

particular crop type.

So, in India, we

are able to do this.

In Europe, we are

able to do this.

In China, unfortunately,

we are not able to

do this for a

variety of reasons.

If I was to show you that same map in a more simple manner,

you could also take that

map and simply turn it

into irrigated versus

non-irrigated agriculture.

And because India is the

number two global irrigator

in the world, and those

are the places in India

where you can see irrigated

croplands in green,

and non-irrigated

croplands in brown or red.

So, again in places

where you have this data,

ground data, we are

able to map them,

but that's not always possible

everywhere on the planet.

We have examples from

Bangladesh as well.

So this is an example

of different rice

types from Bangladesh,

and we had partners

in Bangladesh that

gave us great support.

So one of the jokes that I

have with people that I meet,

and maybe I can do that

with you, is that if

you travel somewhere,

tell us what kind of

crops you're seeing.

I'll give you my email

at the end of this talk,

so that we can utilize that

data to make a better map.

Or if you know people

in other countries.

The last example

I wanna show you,

which I'm really

excited to share is,

this idea of yield assessment.

And when I say yield assessment,

I'm particularly talking

about getting at

actual crop yields

on a field by field basis.

As I indicated earlier,

what if we had thousand

corn and soybean

fields in Dane County,

and what if you were

able to produce yield

data for every one of them?

Now obviously, farmers

who are farming that land,

know exactly what kind

of yields they have.

But that information is in

private hands and government

would like to access it,

or other organizations

would like to access it

for a variety of reasons,

but it's simply not

publicly available.

What if we had the ability,

or produce the ability to

get at these idea of yield data?

And the way that we have been

approaching this problem,

and I think we're

somewhat successful,

but it doesn't work

everywhere, and I want to

show you some of that.

And the way that we

approach this problem is we

look at satellite data.

Here is a landscape with

corn and soy bean data.

So different colors

refer to different

crop types in the

raw satellite data.

And we can take that and

turn into a crop type map.

So green is soybean.

Yellow is corn.

And then, using the

time series data

that I showed you

earlier, we can take the

satellite data and turn it

into a biophysical variable.

And biophysical variable

in this case is a

variable called LAI

or Leaf Area Index,

and is simply defines

how much green leaf

area a particular plot

of land would have.

So if I was to go to a

highly vigor cornfield,

for example, and simply cut

all the leaves of the corn

and lay them down in a

one square meter area,

and if I fill that one

square area one time,

Leaf Area Index would be one.

And if I fill it with

twice, Leaf Area Index

would be two, so

on and so forth.

So with the time series

data, we could convert these

satellite observations

into these biophysical

variables that obviously

have biophysical meaning,

and yield is a biophysical

variable that we

would have to get at.

And we use a simple

formula to do that

and that formula looks

something like this.

So yield is equal to

some Leaf Area Index

times the amount

of solar radiation.

So you can think of

these plants in a

basically a sort

of solar panels.

They have these leaves.

They're absorbing

all the radiation

and they're converting

that solar radiation

through photosynthesis

into biomass.

And at some point, that green

biomass becomes grain yield,

and then you can

allocate part of that

photosynthesis into grain yield.

And the beauty of this

particular approach that

was put forth 40 years ago is

that it nicely lends itself

to remote sensing because one

of these variables is actually,

this particular variable

is the variable

that we can get

from remote sensing.

So, we take these satellite

data for every pixel

or every field in there,

we assemble a time series

and then that time series

variable becomes the

biophysical variable.

And then we can simply

accumulate the amount of

photosynthesis that the

crop is going through,

and then simply

accumulate that and

through some simple conversion,

turn that into a crop yield.

So, here is an example

from Dane County.

So this map is for

corn-planted fields.

So every one of these little

polygons is a corn field,

judged to be corn

field in the year 2003.

So if you're a farmer and

if your field is here,

you can point to it. Right?

And this is the

yield data that we

produced for that

particular year.

So sorry, this is in metric.

So 13 tons would be about

210 bushels per acre.

Pretty good yields.

So as you can see, there

is this wide variety of

yields that you can

get in a Dane County.

Now obviously, producing

a map like that,

as I said earlier, is easy.

Anybody can produce that,

but how good is it? Right?

It's an important question.

So, one of the ways that we can

assess how good that map is,

is to compare it to what the government USDA estimates are.

So, because we have the

entire population of

every field in Dane County,

we can produce histograms.

I love histograms. Right?

So, we can produce

histogram and that

histogram would look

something like that, it

has an interesting shape.

Somewhat bell shaped,

but it's not exactly.

And then, on top of it, we

could plot the USDA estimate

for that year, which is

this red bar over here.

So what this particular

analysis is telling me is

that we are somewhat

over estimating yields

on the average value,

but it's not that bad.

So I think it's a good start.

So one of things that I'm

really excited about doing is to

expand this kind of technology

to the entire Midwest.

Imagine doing this for entire

southern Wisconsin where

croplands are important,

or Illinois or Iowa

for that matter.

And then, produce that

information and then

perhaps compare them to

government estimates,

and then somehow use

that information for

a variety of ways,

including environmental

assessment, so on and so forth.

So, what I wanna do

is conclude by saying

agriculture has

great data needs.

It's important, but it

also has data needs,

and I think remote

sensing can address some

of these needs, but

not all of them.

And it might be able to do

some of these data needs really

well in some locations, but

not well in other locations.

And there are challenges

remain, but I think we're

overcoming them because of

computation power that we have,

and then there are a lot

of exciting developments

that offer a number

of solutions.

And I think we're

doing our part at this

University to address

some of those challenges.

So, here's my email

address as I promised.

So if you travel somewhere

and see interesting crops,

let us know what you're seeing.

But we need to know

where they are.

With that, I'll thank

you and take questions.

[audience clapping]

Share this page