Use of SPOT imagery to perform agricultural land use classification and cropland degradation identification in Kazakhstan and Uzbekistan

munchsistersAI and Robotics

Oct 17, 2013 (3 years and 5 months ago)

74 views






Use of SPOT imagery to perform agricultural land use
classification and
crop
land degradation
identification

in Kazakhstan and Uzbekistan


2
nd

Interim Report to Planet Action and Spot Image








Summary

In the
2
nd

project
year
the
SPOT images acquired via
Planet Action

and
Spot
Image

for the Kyzl
-
Orda
study area
(
Kazakhstan
)

and

Karakalpakstan
(U
zbekistan
)

were used to assess (i) the recent agricultural land use
and
cropping pattern,
and (ii) to identify cropland degradation, re
spectively
.
The

SPOT
images were used to delineate agricultural fields
(see 1
st

interim report). In
the 2
nd

year a

classification routines
was developed
to quantify the agricultural
land use in the study areas
at the field level
(e.g. area estimation, accu
rate crop

identification).
The crop classification routine is based on a classifier ensemble
strategy that fuses the results from Random Forest and Support Vector Machine
algorithms. This method holds great potential for operational land use
monitoring at
the field level, as the field boundaries that were extracted using the
SPOT images can theoretically be used in the next decades (because of the
relative stability of the field borders).
In a pilot study
that was
accomplished
in
the Karakalpakstan region

(
Uzbekistan)

the potential of SPOT images for
identification of degraded land using spectral mixture analysis was shown.

This
method can extend existing
object
-
based
approaches for
identification of
cropland degradation in the large irrigated agricultural
areas

in Middle Asia.

Project Calendar 2
012

Jan
-
Nov



Implementation and testing phase of the classification approach, final
results available (Fabian Löw).

Spring season



Travel to Uzbekistan by Olena Dubovyk

for field sampling
(Karakalpakstan region).

Jan
-
Dec



Implementation and testing phase of a methodology for cropland
degradation identification, first results available (
Olena Dubovyk
).

Satellite Image Analysis and Findings

The results from the project
in 2012
are summarized in way of two
comprehens
ive reports
(
attached below
)
.




Per
-
field crop classification in irrigated agricultural regions

in
M
iddle Asia
using decision fusion
1


Fabian Löw
2
a
, Gunther Schorcht
a
,

Ulrich Michel
b
,

Stefan Dech
a,
c
,

Christopher Conrad
a


a
Department of Remote Sensing, Julius
-
Maximilians
-
Universität Würzburg, Am Hubland, 97074
Würzburg, Germany;
b
Department of Geography, University of Education, Czernyring 22/ 11
-
12,
69115 Heidelberg, Germany;
c
German Remote Sensing Data Center (DFD), German

Aerospace
Center (DLR), Oberpfaffenhofen, 82234 Wessling, Germany

ABSTRACT

Accurate crop identification and crop area estimation are important for studies on irrigated agricultural systems, yield
and water demand modeling, and agrarian policy developmen
t. In this study a novel combination of Random Forest (RF)
and Support Vector Machine (SVM) classifiers is presented that (i) enhances crop classification accuracy and (ii)
provides spatial information on map uncertainty. The methodology was implemented ov
er four distinct irrigated sites in
Middle Asia using RapidEye time series data. The RF feature importance statistics was used as feature
-
selection strategy
for the SVM to assess possible
negative effects

on

classification accuracy caused by an oversized f
eature space. The
results of the individual RF and SVM classifications were combined
with rules
based on posterior classification
probability and estimates of classification probability entropy. SVM classification performance was increased by feature
selec
tion through RF. Further experimental results indicate that the hybrid classifier improves overall classification
accuracy in comparison to the single classifiers as well as
user´s and producer´s accuracy.


Keywords:

Ensemble classifier,
f
eature
selection
,

Hughes phenomenon, map uncertainty, random forest (RF),
R
apidEye, support vector machine (SVM
)


1.1

Study sites

Four test sites in Middle Asia were chosen: Karakalpakstan (KKP), Khorezm (KHO), Kyzl
-
Orda (KYZ), and Fergana
Valley (FER). They are located among
the Amudarya and Syrdarya rivers in Middle Asia and are characterized by vast
agricultural systems. Climate is arid continental and dry with less than 100 mm precipitation per year, thus agriculture is
limited to irrigated lands.
Crops found in the study s
ites are: rice, winter wheat, cotton, alfalfa, melons, sorghum, maize,
and fruit trees. Further rotations or winter wheat and other crops (e.g. rice or sorghum) can be found.
The topography of
the sites is flat, with average elevations of 66
m

(KKP), 85
m

(K
HO), 128
m

(KYZ), and 577m

a.s.l. (FER). The median
field sizes are 1.71 ha (KKP), 2.14 ha (KYZ), 3.21 ha (KHO), and 5.47 ha (FER). In this study, a subset of 40x40km was
chosen in each site (
Figure 1
).




1

Published as: F. Löw, G. Schorcht, U.
Michel, S. Dech, and C. Conrad, “Per
-
field crop classification in irrigated agricultural regions in middle Asia
using random forest and support vector machine ensemble,” in Proceedings of SPIE 8538, Earth Resources and Environmental Remo
te Sensing/GIS
Appl
ications III, 2012.

2

Contact: fabian.loew@uni
-
wuerzburg.de, Tel.: +49 (0)931 / 31
-
84072


.

Figure
1
: Satellite image
ry for four sample sites in the main irrigated areas in Central Asia, showing a variety of different
field sizes, shapes, and cropping pattern. The imagery is displayed using a 5
-
2
-
1
-
band combination of the RapidEye sensor.
Contrast is adjusted to each ban
d separately. Scenes are from the summer season (June
-
August) 2011.

1.2

Satellite Images

Multispectral images from the RapidEye system were acquired over the test sites in 2011 and 2009 (KHO).
The
RapidEye sensor has a nominal ground sampling distance (GSD) of

6.5m, with five spectral bands:
Blue (440
-
510 nm),
Green (520
-
590 nm), Red (630
-
685 nm), RedEdge (690
-
730 nm), and
near infrared (
NIR
)

(760
-
850 nm).
For each site,
eight images were acquired that cover the vegetation period from April


October. The Rapid
Eye images were co
-
registered to high
-
resolution SPOT
-
5 images (2.5m) in KYZ and KHO, and with the use of ground control point (GCP)
identification in the images in KKP and FER. A nearest
-
neighbor transformation of all images to the Universal Trans
Mercato
r
(UTM)
reference projection (WGS84 datum) resulted in sub
-
pixel accuracies (RMSE


0.9 pixels) for all
images. Atmospheric correction was performed in
the
ATCOR3 module, based on the M
ODTRAN radiative transfer
code.

1.3

Sampling data and field vector database

Reference samples for training and validation purposes were collected during extensive field surveys in 2011 (2009 in
KHO).
In FER, KHO and KKP a two
-
time sampling was implemented to distinguish rotations of winter wheat with
summer crops (1
st

sampling
between March and June and 2
nd

between July and September
).

The sampling sites
comprised fields of rice, cotton, melons, winter wheat, rotations of wheat and other crops, alfalfa, sorghum, maize,
sunflower, fruit trees, and fallow lands.

The total number o
f field samples was 421 (KHO), 507
(KKP), 409 (FER) and
405 (KYZ).

The agricultural fields were
derived

using object
-
based image segmentation
.

The multi
-
resolution and spectral difference
segmentation algorithms included in the commercial software Definiens

Developer 8 (Definiens AG, Munich,
Germany)
(Trimble Germany GmbH, 2011)

were

used.

Fo
r KHO and KYZ 2.5m SPOT
-
5 data
provided by Planet Action
and
recorded in spring of 2006 and 2010, respectively, were used for segmentation. Fields for KKP and FER were
derived

with RapidEye 6.5m images

from spring May 2011
, where the five spectral bands an
d a variance texture image
were used as input.

A comprehensive set of spectral and semivariogram derived features were extracted for each field and time step.

The
features

can be categorized in five groups: in the first group the reflectance information from the five RapidEye bands is
contained. Vegetation indices belong to the second group.
S
everal VIs were calculated based on the information of the
RedEdge channel and are

composed to the third group of input features.
The fourth group comprises features describing
the curve properties
between Red and NIR
(Conrad et al., 2012)

utilizing the angle and Euclidian distances between Red,
RedEdge and NIR bands
.
The fifth group consists of features derived from each objects semivariogram. The
semivariogram quantifies the spatial associations of the values of a variable, and measures the degree of spatial
correlation between different pixels in an object
(Ruiz, Recio, Fernández
-
Sarría, & Hermosilla, 2011)
. From each
semivariogram 11 features were calculated using the software FETEX 2.0
(Ruiz et al., 2011)
, the features are fully
described in
(Balaguer, Ruiz, Hermosilla, & Recio, 2010)
.

A two level classification scheme was implemented. In the first level the accuracy of
the individual classifiers (Random
Forest and Support Vector Machine) was boosted by optimally configuring the algorithms (e.g. number of trees in the
Random Forest ensemble). In the second level the results were fused at the decision level by using local
uncertainty
information from the algorithms
(Löw, Schorcht, Michel, Dech, & Conrad, 2012)
. The final decision was done by
selecting the class with lowest uncertainty.


2.

RESULTS AND DISCUSSI
ON

2.1

Accuracy

The OA

of the ensemble range between 93.6
0
%

and 95.
1
0
%,
which

indicates

its

overall
high
classification performance
(
Table 1
). Further the ensemble strategy outperforms the RF and the
SVM

in

all test sites
.

The RF achieved OA between
89.46 and 94.33%, the regular SVM between 86.70% and 92.30
%. The McNemar test shows

that the results achieved
with the ensemble approach significantly outperform most of its individual regular classifiers at the 0.05 confidence
level
, with some exceptions
.
Using a traditional maximum likelihood classifier (MLC) resulted in OA between 57.
13%
and 70.51%. MLC was outperformed by all non
-
parametric approaches.
The absolute difference in OA between the
classifiers over the study sites was lo
west for the ensemble with 1.91%
, 4.87% for RF,
5.30% for SVM, and 16.55% for
MLC
.

This indicates a better stability of the results from the ensemble concerning classification accuracy, irrespective of
the number of classes.

Table
1
: Overall accuracies [%] using the individual classifiers (RF, SVM), the SVM after feature selection (SVM
selected)
and the ensemble (ENS). Significant difference (α = 0.05) of the ensemble in comparison to
the other classifiers
was
tested with the McNemar test and is
indicated with an asterisk (*).


Study site


Method

KHO

KKP

KYZ

FER

Absolute
difference of
OA

MLC

57.13

53.96

66.74

70.51

16.55

RF

89.46

88.09

92.78

94.33

4.87

SVM

86.70

87.70

92.30

92.00

5.30

E
nsemble

93.60

94.70

94.78

95.51

1.91


The producer´s and user´s accuracies indicate that the ensemble performs more accurate in most of the cases

(Figure
2
).

While some classes show almost no differences, such as “Rice” in KYZ or “
Winter Wheat
” in KHO, the differences tend
toward higher accuracies in most of the classes. In KYZ the ensemble performed significantly better in allocating the
class “
Wi
nter Wheat
” correctly. Altogether the overall improvement is most pronounced in KHO, KYZ, and FER. It is
less pronounced in KKP, where few classes tend towards the negative, e.g. “Cotton” or “Sorghum/Maize”.







Figure
2
:

Differences

between user´s and producer´s accuracy [%] achieved by classifier ensemble and the individual classifiers in (a)
KHO, (b) KKP, (c) KYZ, and (d) FER.

3.

CONCLUSIONS

Combining RF and SVM in an ensemble classifier provided more accurate results over all test
sites.

The
McNemar test
indicated a significant improvement in classification accuracy of the ensemble compared to RF and SVM in two of the
four sites.
In contrast

to the standalone methods the accuracy of the ensemble was almost equal high in all sites, w
hich
shows that the ensemble strategy can successfully be transferred onto quite different agricultural landscapes in Middle
Asia.

The highest gain in terms of O
A was
achieved

at level one in the ensemble strategy by selecting optimal
parameters for the RF

(number of trees) and the SVM (features selection).

At level two the highest gain was achieved in
producer´s and user´s accuracies, which were observed to increase in most of the cases.

The methodology provides
spatial informatio
n on classification uncert
ainty
:

i
t was shown that correctly classified cases are correlated wit
h high PPE
and low RCPE values
(Löw et al., 2012)
.

ACKNOWLEGMENTS

We thank the German Aerospace Agency (DLR) for providing data from the RapidEye Science Archive (RESA), and the

Deutsche Gesellschaft für International
e Zusammenarbeit


(
GIZ
)

for logistical support of the field surveys

in
Uzbekistan and Kazakhstan
. SPOT 5 images were made available through the S
pot Image project Planet Action.

-20
-10
0
10
20
30
40
50
cotton
wheat_other
trees_crops
wheat_fallow
rice
sorghum_maize
fallow
a)

-20
-10
0
10
20
30
40
50
cotton
fallow
rice
sorghum_maize
wheat_fallow
alfalfa_1y
melons
wheat_other
b)

-20
-10
0
10
20
30
40
50
fallow
rice
wheat_fallow
alfalfa_1y
alfalfa_3y
c)

-20
-10
0
10
20
30
40
50
cotton
fallow
fruittree
wheat_fallow
wheat-other
d)

LITERATURE

Balaguer, A., Ruiz, L.
, Hermosilla, T., & Recio, J. (2010). Definition of a comprehensive set of texture semivariogram
features and their evaluation for object
-
oriented image classification.
Computers & Geosciences
,
36
(2), 231

240.
doi:10.1016/j.cageo.2009.05.003

Conrad, C., Fritsch, S., Lex, S., Löw, F., Rücker, G., Schorcht, G., Sultanov, M., et al.
( 0 ). Potenziale des Red Edge
Kanals von RapidEye zur nterscheidung und zum Monitoring landwirtschaftlicher Anbaufr chte am eispiel des
usbekischen Bewaesserun
gssystems Khorezm. In E. Borg, H. Daedelow, & A. R. Johnson (Eds.),
RapidEye
Science Archive (RESA)
-

Vom Algorithmus zum Produkt

(pp. 201

217). Berlin:GITO: Springer.

Löw, F., Schorcht, G., Michel, U., Dech, S., & Conrad, C. (2012).
Per
-
field crop classif
ication in irrigated agricultural
regions in middle Asia using random forest and support vector machine ensemble.
Proceedings of SPIE 8538,
Earth Resources and Environmental Remote Sensing/GIS Applications III
. Edinburgh. doi:doi: 10.1117/12.974588

Ruiz, L
., Recio, J., Fernández
-
Sarría, A., & Hermosilla, T. (2011). A feature extraction software tool for agricultural
object
-
based image analysis.
Computers and Electronics in Agriculture Electronics in Agriculture
,
76
(2), 284

296.
Retrieved from http://www.sci
encedirect.com/science/article/pii/S0168169911000573

Trimble Germany GmbH. (2011).
eCognition Developer 8.7 Reference Book
.





Object
-
based cropland degradation identification: a case
study
in
Uzbekistan



Olena Dubovyk
a
,

Alexander Lee
b
,
Murod
Sultanov
b
,

Asia Khamzina
a

a
Center for Development Research, University of Bonn, Walter
-
Flex Str. 3, 53113 Bonn, Germany
;

b
Khorezm Rural Advisory Support Service (KRASS), Khamid Alimjan str.,14, 220100 Urgench,
Uzbekistan

ABSTRACT

Sustainability of irriga
ted agriculture
-
based economies,

such as in Central Asia
,

is threatened by cropland degradation.
The field
-
based identification of the degraded agricultural areas
can aid in

developing
appropriate
land rehabilitation and
monitoring programs.
This
report is

focused on
the object
-
based
change detection

and spectral mixture analysis to
develop an approach for identifying parcels of irrigated degraded cropland in Northern Uzbekistan, Central Asia. A
linear spectral unmixing

i
s applied to the multiple
SPOT

i
mage
s
. Considering a spectral dimensionality of
SPOT 5, a
multiple 3
-
endmember
model
(green vegetation, water,
and soil
) was set up for the analysis. The spectral unmixing
results were valid, as indicated by the
low values of
ove
rall root mean square errors in

a range below <2.5%
.
The
following analysis will include

the object
-
based
change detection
to reveal the
cropland
,
af
fected by
the
degradation
processes

to varying degrees. The proposed approach could be
elaborated

for a field
-
based moni
toring of cropland

degradation
in similar landscapes of Central Asia and elsewhere.


Keywords:

Land

cover

degradation, object
-
based analysis, spectral unmixing, change
detection
,
SPOT 5
, irrigated
cropland
, Central Asia.

1.

INTRODUCTION


In Central Asia
, about 22 million peopl
e depend on irrigated agriculture, while over 50% of irrigated cropland are
reportedly degraded, causing agricultural production losses estimated at 2
billion USD per year.

The
problem

of
cropland
degradation originated

during the Soviet Union

time
, when
t
raditional
agricultural activities were extended through
the
construction of
the large
-
scale
irrigation system for cultivation of cotton crop, the white gold’ of Central Asia.
T
he total
irrigated area
was
increased by
70%
to 9.4 × 106 ha in 1989

(Lewis, 1
962)
.

The
decades of water and land
mismanagement have led to the severe environmental problems, including local
and regional
climate change due to the
Aral Sea shrinkage, degradation of land and water resources, biodiversity loss, and air, soil and wat
er pollution.
Despite
the

large scale of the described problem,
scarce

information on
spatial extent and severity of
land degradation is
available
in Central Asia

(Dubovyk et al., 2012)
.

In Uzbek
istan,
which is a downstream and

double land
-
locked country
, the

irrigated agriculture compris
es 22
% of
the
GDP, 60% of
the
foreign exchange receipts, an
d 33
% of
the
employment
.
The national agricultural sector depends totally
on irrigation water supply du
e to the aridity of the regional climate.
According to the Intergovernmental Panel on
Climate Change, the existing water stress in Central Asia is more likely to aggravate due to climate change
, posing new
challenges on the agricultural sector

(IPCC, November 12, 2011)
.
E
laboration of
sustainable
practices for natural
resource

management

in
response to land degradation and water scarcity

requires

accurate information on the
crop
land’s
state.

Among different methods
for monitoring land degradation, remote sensing provides a cost
-
effective evaluation over
extensive areas
(Prince et al., 2009)
, whereas in
-
situ process studies are resource demanding, and thus are usuall
y
conducted at the plot

or small
-
catchment scale. Remote sensing based approaches of land degr
adation assessment are
especially useful in developing countries, where reliable data
and financial
means
are

limited.

M
ost of the

remote
sensing applications
have
deal
t

with
a
direct observation of visible features

of degraded areas
, such as different for
ms of
soil erosion

(Li et al., 2007)
.
Many

studies focused on natural and semi
-
natural environments

(Röder et al., 2008)
,
while
less
attention was paid to monitor

cropland degradation
.
The common approach for land degradation analysis with
remote sensing methods considers spatio
-
temporal dynamics of land us
e and land cover (LULC) changes. This approach
often implies image clas
sification

(Gao and Liu, 2008)
, eventually followed by change detection to quantify chang
es in
LULC classes
,

(e.g., Chen and Rao, 2008)
.
Another technique analy
zes

gradual changes in
the
vegetation
cover

over
time, using time series of satellite imagery
(e.g., Röder et al., 2008)
.

Trend analys
is of

time series
has

been recognized
as
an effective means for

land degradation
monitoring

on regional and global scale

(Fensholt and Proud, 2012)
.
A more

detailed
assessment

based on medium spatial resolution satellite time series

is
, however,

restricted in some regions of the
world, such as in Central Asia
,
by the
low
availability of
the imagery

that

cover
s

a
sufficiently long time period
.

The field
-
based identification of degraded agricultural areas
can aid in

developing land rehabilitation and monitoring
programs. In
arid and semi
-
arid environments
, traditional pi
xel
-
based analysis of satellite data could lead to inaccurate
mapping of degraded cropland due to
the problem of
mixed pixels.

In contrast, object
-
based image analysis can handle
these issues, at the same time providing parcel
-
specific information on the l
and’s state

(Pena
-
Barragan et al., 2011)
.
Assessment of vegetation cover de
cline, as a main indicator of land degradation, could be challenging with
conventionally
-
applied vegetation indices, due to, for example, their sensitivity to soil background

(Huete et al., 2002)
.
Spectral mixture analysis (SMA), which provides an accurate quantitative estimation of proportional vegetation

cover at
a subpixel level is
,

therefore
,

suitable for monitoring vegetation changes in
arid

landscapes
(Tromp and Epema, 1998)
.

In this light,
this study aimed

to
provide
field
-
specific
information about land degradation

for a case study in Uzbekistan
by combining the

object
-
based

change detection and SMA
.

2.

STUDY AREA

The study was conducted in the
lower reaches

of the Amu D
arya River in

the Khorezm R
egion
of Uzbekistan
and in

the
southern

part

of
Autonom
ous Republic of Karakalpakstan

(SKKP)

in Uzbekistan

(Figure 1)
.
This area belongs to the
Central Asian semi
-
desert zone
with

the

arid and
extreme continental climate.

Mean a
nnual precipitation of
ca.

100 mm
falls mostly

outside the growing season
(April
-
October),
while
mean annual evaporation is ca.
1500 mm
(Glazirin et al.,
1999)
. Crop cultivation depends solely on irrigati
on water supply, supported by the dense irrigation network
(Conrad et
al., 2010)
.

Most of the cropland is occupied by cotton (60
-
70%) and winter whe
at (20
-
30%)
(Shi et al., 2007)
. These
crops

are
cultivated under the

state pr
ocurement system which assigns

production targets

and predefines spatial cr
opping
patterns

for these strategic crops
.

On the remaining
crop
land, farmers grow sorghum, maize, melons, watermelons,
vegetables
, and horticultural crops
.



Figure
2
.
Location of the study area in

Uzbekistan

(left) and in Central Asia (right).
The

study area’s

border coincides with
the extent of irrigated
crop
land
.

A
bout 20% of the
agricultural

land

is

classified
as

low
suitable for crop

cultivation
,

according to the local soil quality
measure
, called soil
bon
itation

(Khamzina et al., 2008)
. Irrigation water is d
istributed through the
dense
irrigation
can
al

network
. The irrigation canals and drains are usually not properly maintained

which results in 40% of water

losses from
evaporation and
filtration
(Conrad et al., 2007)
.
Reportedly, groundwater table

level
s within

the range of 1.5
-
1.6 m occur
on over 80% o
f the irrigated area, imposing the risk for
secondary
soil salinization and waterlogging

(Ibrakhimov et al.,
2007)
.
A combination of

shallow groundwater table,
high

evapotranspi
ration and inadequate drainage leads to salt
accumulation in the surface soil, altering soil physicochemical properties and ultimately triggering
land
degradation
processes
.

3.

DATA AND METHODS

The data used

in the study

included (i) raster data:
SPOT 5

image
s
and (ii) vector data:
boundaries of agricultural
field
s,

digitized from
the
cadastral maps
.

The
vector data

w
as

collected from
the
database of
ZEF/UNESCO
project
(
http://www.khorezm.zef.de/
)
.
Field survey

and e
xpert

interviews were held with
irrigation engineers, land managers and
cadastral officers

in
2010

to
get more insights on cropland degradation and
abandonment

in the study region.

3.1

Satellite data and preprocessing

The analyses were based on the
SPOT

5 TM i
mages
, recorded in
1998, 2008, and 2010
. F
or these years
,

the cloud
-
free

images were available
for the crop growing season
.

All images were
geometrically adjusted to
a 2.5 m SPOT
-
5 scene
,
projected to the
UTM coordinate system

(
zone 41
)
, based on

the

differential
GPS

points

(Conrad

et al., 2010)
.

The
n
earest neighbor resampling with a
second

degree polynomial model resulted in an overall positional error <0.5 pixel.
The

entire dataset consisted
of
6th images
.

For the subsequent analysis, the
field
mask was applied to the images,
comprising

the boundaries of the agricultural
fields
.
.A
ll non
-
agricultural areas

were excluded from the analysis
.

3.2

Spectral mixture analysis

As this study was aimed at determining
the

vegetation cover for agricultural fields,

SMA

was employed to quantify
abundance fractions

of

land cover
at a subpixel level

(Smith et al., 1990)
. Spectral mixture theory assumes that the
reflectance of a pixel is a mixture of the reflectance of covers or endmembers, contained within the pixel.
Our

study

impl
emented

the image
-
based approach

where

the derived endmembers represent
ed

the spectra, measured at the same
scale as the satellite data
used

(Lu et al., 2007)
.

Consequently, the endmembers were selected using a pixel purity index,
which was computed, based on the minimum noise fraction (MNF) transformation

(Green et al., 1988)
, applied to
four
bands of
the

August 2010

SPOT

image
.

In the linear
of spectral unmixing
approach, the mixed spectrum is expressed in terms of the linear combination of the
spectra of the pure components, based on their fractional area

(Smith et al., 1990)
.
Considering the recent successful
application of linear mixtu
re approach for land degradation studies

(Röder et al., 2
008)
, a linear constrained unmixing
model was applied to derive the subpixel fractions according to:




R
i

=


(1)

and




(Vlek)

where

R
i

=
reflectanc
e of the mixed spectrum in band
i

R
E
ij
=

reflectance of the endmember spectrum
j

in band
i

F
j

=

a
fraction of endmember
j

n=

number of spectral endmembers



=

a residual error in band
i

j
= 1
-
n

i
= 1
-
6

To assess
a
fit of the spectral unmixing model, the root mean square error (RMSE) was calculated
for all images, and t
he
histograms of the derived fractions as well as their
residual images

were
visually
analyzed
.

4.

RESULTS


4.1

Spectral Mixture Analysis

A 3
-
endmember model
, including
green vegetation

(GV)
, soil

(
S)
, and water

(WT)

was set up
.

After determination of
image endmembers, their relationship to elements of the landscape was established

(Figure
2 and 3
)
.

G
V consisted of all
vegetation

types
within the fields.
S
rep
resented sand
y
soils and

other
soils, respectively.

WT referred mainly to

shallow
bright water surfaces within the flooded rice fields.

The effect of shadow was not considered
,

as it is minor for the sparse
canopies in arid and semiarid environments

(Trodd and Dougill, 1998)
.


Figure 1. False
-
color composite of the SMA results, base
d on August 2010
SPOT
-
5 image. On the map, red color
represents
soil fraction, green is vegetation fraction,
blue stands for water fraction, and white color shows the masked out
areas.

A quant
itative assessment of fraction
images is complicated in heteroge
neous arid landscapes. A combination of
different validation approaches is,

therefore
, needed to evaluate results of SMA. The statistical validity of the
spectral
unmixing

was confirmed by
the low values of RMSE in a range below 2.5%.

T
he histograms of the

individual
abundance fractions did
not generally exceed the values from 0 to

1, and the residual bands showed no systematic
patterns.

The similar results were reported in the land degradation studies by
Chen and Rao
(2008)

and Röder et
al.
(2008)
,
suggesting overall plausibility of SMA results.





Figure 2.
The elements of the landscapes, represented in
SMA: green vegetation

(upper left), bare soil (upper right)
and shallow ater surfaces (lower left). Photos are taken by
O. Dubovyk in summer 2012.


5.

FURTHER ANALYSIS

The set of endmembers, selected on the August 2010 image, will be applied to the rest of the images. The usage
of the
same endmembers for analysis of multitemporal images improves the accuracy of change detection, as it allows for
direct comparison of calculated fractional covers
(Elmore et al., 2
000)
.

T
h
e following step

will include change detection between the images 0f 1998, 2006, 2008 and 2010
to identify degraded
agricultural fields. The differences will be computed using a change vector analysis (CVA)
(Malila, 1980)
. CVA allow
s
determining the magnitude and direction of change between two time steps.

6.

CONCLUSIONS

This
report

presented an approach for
object
-
based identification of
cropland degradation
in the larg
e irrigated
agricultural areas
on the example of the study region
in
Uzbekistan.

The analysis involves

spectral linear unmixing,
follow
ing

by

the object
-
based change
detection

and classification
of the detected changes
per se
. The results

will reveal

spatial distribution of degraded
agricultural
fields

based on the difference
between
the analyzed
years
.
These
fields
should
be the main target of
land
rehabilitation

and mitigation

measures.
.

REFERENCES

CHEN, S. & RAO, P. 2008. Land degradation monitoring using multi
-
temporal Landsat TM/E
TM
data in a transition zone between grassland and cropland of northeast China.
International
Journal of Remote Sensing,

29
,

2055
-
2073.

CONRAD, C., DECH, S., HAFEEZ, M., LAMERS, J., MARTIUS, C. & STRUNZ, G. 2007.
Mapping and assessing water use in a Centra
l Asian irrigation system by utilizing MODIS
remote sensing products.
Irrigation and Drainage Systems,

21
,

197
-
218.

CONRAD, C., FRITSCH, S., ZEIDLER, J., RÜCKER, G. & DECH, S. 2010.
Per
-
Field Irrigated
Crop Classification in Arid Central Asia Using SPOT and ASTER Data.
Remote Sensing,

2
,

1035
-
1056.

DUBOVYK, O., MENZ, G. & KHAMZINA, A. Year. Trend analysis of MODIS time
-
series using
different vegetation indices for monitoring of cropla
nd degradation and abandonment in
Central Asia.
In:

IEEE International Geoscience and Remote Sensing Symposium, 22
-
27
July 2012 Munich.

ELMORE, A. J., MUSTARD, J. F., MANNING, S. J. &

LOBELL, D. B. 2000. Quantifying
Vegetation Change in Semiarid Environments: Precision and Accuracy of Spectral Mixture
Analysis and the Normalized Difference Vegetation Index.
Remote Sensing of Environment,

73
,

87
-
102.

FENSHOLT, R. & PROUD, S. R. 2012. Ev
aluation of Earth Observation based global long term
vegetation trends
-

Comparing GIMMS and MODIS global NDVI time series.
Remote
Sensing of Environment,

119
,

131
-
147.

GAO, J. & LIU, Y. 2008. Mapping of land degradation from space: a comparative study of
Landsat
ETM+ and ASTER data.
International Journal of Remote Sensing,

29
,

4029
-
4043.

GLAZIRIN, G. E., CHANISHEV, S. G. & CHUB, V. E. 1999. Brief outlines of climate in
Uzbekistan. Tashkent: Chinor ENK (in Russian).

GREEN, A. A., BERMAN, M., SWITZER, P. & C
RAIG, M. D. 1988. A transformation for ordering
multispectral data in terms of image quality with implications for noise removal.
IEEE
Transactions on Geoscience and Remote Sensing,

26
,

65
-
74.

HUETE, A., DIDAN, K., MIURA, T., RODRIGUEZ, E. P., GAO, X. & FE
RREIRA, L. G. 2002.
Overview of the radiometric and biophysical performance of the MODIS vegetation indices.
Remote Sensing of Environment,

83
,

195
-
213.

IBRAKHIMOV, M., KHAMZINA, A., FORKUTSA, I., PALUASHEVA, G., LAMERS, J. P. A.,
TISCHBEIN, B., VLEK, P. L
. G. & MARTIUS, C. 2007. Groundwater table and salinity:
Spatial and temporal distribution and influence on soil salinization in Khorezm region
(Uzbekistan, Aral Sea Basin).
Irrigation and Drainage Systems,

21
,

219
-
236.

IPCC. November 12, 2011.
Assessment
Report
[Online]. Available:
http://www.ipcc.ch/ipccreports/ar4
-
wg1.htm

[Accessed].

KHAMZINA, A., LAMERS, J. P. A. & VLEK, P. L. G. 2008. Tree establishment under deficit
irrigation on degraded agri
cultural land in the lower Amu Darya River region, Aral Sea
Basin.
Forest Ecology and Management,

255
,

168
-
178.

LEWIS, R. 1962. The irrigation potential of Soviet Central Asia.
Annals of the Association of
American Geographers,

52
,

99
-
114.

LI, X., WANG, Z.
, SONG, K., ZHANG, B., LIU, D. & GUO, Z. 2007. Assessment for Salinized
Wasteland Expansion and Land Use Change Using GIS and Remote Sensing in the West Part
of Northeast China.
Environmental Monitoring and Assessment,

131
,

421
-
437.

LU, D., BATISTELLA, M., MAUSEL, P. & MORAN, E. 2007. Mapping and monitoring land
degradation risks in the Western Brazilian Amazon using multitemporal Landsat TM/ETM+
images.
Land Degradation & Development,

18
,

41
-
54.

MALILA, W. A. 1980. Change vector anal
ysis: an approach for detecting forest changes with
Landsat
Proc. 6th Annual Symposium on Machine Processing of Remotely Sensed Data
,

326
-
335.

PENA
-
BARRAGAN, J., NGUGI, M. K., PLANT, R. E. & SIX, J. 2011. Object
-
based crop
identification using multiple veg
etation indices, textural features and crop phenology.
Remote
Sensing of Environment,

115
,

1301
-
1316.

PRINCE, S. D., BECKER
-
RESHEF, I. & RISHMAWI, K. 2009. Detection and mapping of long
-
term land degradation using local net production scaling: Application
to Zimbabwe.
Remote
Sensing of Environment,

113
,

1046
-
1057.

RÖDER, A., UDELHOVEN, T., HILL, J., DEL BARRIO, G. & TSIOURLIS, G. 2008. Trend
analysis of Landsat
-
TM and
-
ETM+ imagery to monitor grazing impact in a rangeland
ecosystem in Northern Greece.
Remot
e sensing of environment
112
,

2863
-
2875.

SHI, Z., RUECKER, G., MUELLER, M., CONRAD, C., IBRAGIMOV, N., LAMERS, J.,
MARTIUS, C., STRUNZ, G., DECH, S. & VLEK, P. 2007. Modeling of cotton yields in the
amu darya river floodplains of Uzbekistan integrating mul
titemporal remote sensing and
minimum field data.
Agronomy Journal
99
,

1317
-
1326.

SMITH, M. O., USTIN, S. L., ADAMS, J. B. & GILLESPIE, A. R. 1990. Vegetation in deserts: A
regional measure of abundance from multispectral images.
Remote Sensing of Environm
ent,

31
,

1
-
26.

TRODD, N. M. & DOUGILL, A. J. 1998. Monitoring vegetation dynamics in semi
-
arid African
rangelands: Use and limitations of Earth observation data to characterize vegetation structure.
Applied Geography,

18
,

315
-
330.

TROMP, M. & EPEMA, G. F.
1998. Spectral mixture analysis for mapping land degradation in
semi
-
arid areas.
Geologie en Mijnbouw,

77
,

153
-
160.

VLEK, P., LE, Q., & TAMENE, L. (2008). LAND DECLINE IN LAND
-
RICH AFRICA: A
CREEPING DISASTER IN THE MAKING. CGIAR SCIENCE COUNCIL
SECRETARIA
T. ROME: ITALY, 62 PP.






(Further) Project Goals

2013

a)

To investigate
and quantify
patterns of
crop
land degradation in the
pilot
region,

b)

Providing recommendations to the local policy
-
makers on sustainable
wat
er and land resource management, and
agricultural monitoring, based

on remote sensing technologies,

c)

Provide a final and extended report of the project (by the end of 2013).