Performance of an Object-Based Neural Network Classifier on Land Cover Characterization in Amazon, Brazil

chickenchairwomanAI and Robotics

Oct 19, 2013 (3 years and 7 months ago)

56 views


Future



Discussion


Introduction


Methodology


Results


Abstract

There

are

three

types

of

data

used

in

the

project
.

They

are

IKONOS,

ASTER,

and

Landsat

TM,

representing

high

to

low

spatial

resolution

between

4

meters

and

30

meters
.

The

acquisition

date

of

the

IKONOS

data

is

October

14
,

2000
.

The

four

bands

used

are

blue

(
0
.
45
-
0
.
53

μm),

green

(
0
.
52
-
0
.
61

μm),

red

(
0
.
64
-
0
.
72

μm

)

and

near

infrared

(
0
.
77
-
0
.
88

μm

)

at

4
-
meter

resolution
.

Landsat

TM

data

were

acquired

on

May

26
,

1996
.

The

ASTER

image

was

acquired

on

July

29
,

2000

and

has

three

bands,

two

of

which

are

visible

and

one

of

which

is

near

infrared

bands

at

15
-
meter

spatial

resolution
.

The

three

bands

used

are

0
.
52
-
0
.
60

μm
,

0
.
63
-
0
.
69

μm

and

0
.
76
-
0
.
86

μm
.


A

series

of

image
-
preprocessing

operations

were

performed

to

ensure

the

proper

registration

and

the

compatibility

of

the

images
.

Neural

network

classifier

is

the

main

target

to

be

examined
.

A

standard

maximum

likelihood

classifier

was

used

as

reference
.

k

In line with the object
-
oriented approach in the development of the Amazon Information
System, an object
-
based neural network classifier is implemented with the new system
architecture. This research compares the performance of a neural network classifier to that of
a conventional classifier. The project analyzed three images at different
spatial resolutions to
examine the results from the two classifiers on images at different scales. The data subsets
used are from IKONOS (4 meters), ASTER (15 meters), and Landsat TM (30 meters). The
data were acquired in Altamira, Brazil, a typical eastern Amazon tropical area with a collage
of cultivated land, forest, river,and city. A series of pre
-
processing procedures, such as
registration and cloud masking, were applied to assure that the actual subsets cover exactly
the same area. Research results confirm that a neural network classifier, using multiple
source data, yields superior results compared to a maximum likelihood classifier. The object
-
oriented approach to the implementation adds flexibility in interface, interaction, versioning,
and porting. Future studies will focus on the development of a parallel
-
based strategy to
shorten training time and the time for constructing alternative neural networks.

Performance of an Object
-
Based Neural Network Classifier on Land Cover Characterization in Amazon, Brazil


An

object
-
based

neural

network

constructed

upon

the

principle

of

a

multiple

layered

backpropagation

perceptron

was

implemented

in

the

Amazon

Information

System,

with

open

options
.

This

paper

tests

if

the

neural

network

satisfies

the

functional

requirements

and

demonstrates

a

potential

for

superior

classification

capability

compared

to

conventional

digital

image

classifiers,

such

as

the

maximum

likelihood
.

The

following

are

the

main

objectives
:


Testing

the

effect

of

changing

the

number

of

neurons

used,

learning

rate,

and

training

samples,

to

guide

the

optimization

of

the

classifier

design

and

operation
.



Comparing

the

performance

of

a

neural

network

classifier

to

other

classifiers,

e
.
g
.

maximum

likelihood,

to

see

if

the

neural

network

is

superior

in

tropical

land

cover

characterization
.


Examining

the

neural

network

classifier

with

satellite

images

at

multiple

scales,

or

multiple

spatial

resolutions,

in

extracting

land

cover

features
.


A. Hidden Units

Given a set sample, the accuracy of the land
cover characterization changes slightly
with the number of hidden units. In
general, over
-
structured or under
-
structured neural networks show defects.

1 (under
-
structured)

8 (properly
-
structured)

150 (over
-
structured)

B. Hidden Layers

The relationship between hidden layers and the accuracy is
similar to those of between hidden units and the accuracy. In
other words, over
-
structured (over
-
fitting) or under
-
structured
neural networks may occur.

3 (over
-
structured)

2 (slight over
-
structured)

1 (properly in this case)

C. Training Samples

160 pixels

350 pixels

1100 pixels

NN

MLC

NN

MLC

NN

MLC

The following can be noticed from this set of images.


Neural networks (NN) are superior to maximum likelihood classifiers in accurately detecting land cover features. This is espe
cia
lly true when the
training samples are limited. Note the incorrect classification of water surfaces by MLC in the first two cases, in contrast
to
these by NN.


The accuracy of NN classifier has less to do with the number of training samples than with the proper training sample.


NN may be over
-
trained. Feeding correct training samples to neural network classifiers is important for achieving desirable accu
racy.

NN

MLC

NN

MLC

NN

MLC

IKONOS (4m)

ASTER (15m)

Landsat TM (30m)

E. Scale Effects

0.1

0.3

0.5

0.7

0.9

21


23


24


24


37


D. Learning Rate

The neural network system converges faster to
the expected overall error for the network with
a higher learning rate.

Further study with the neural network classifier will focus on better pre
-
processing,
training optimizing, complicated applications and hybrid classifier.


A multiple layered backpropagation/feedforward classifier was implemented with object
-
oriented programming, which gives the flexibility to construct a variety of neural network
architectures. The performance of the classifier was examined internally and externally.
Internally, the classifier is applied with different hidden units, hidden layers, learning rate,
and training samples. Externally, it is compared with a standard maximum likelihood
classifier and with multiple scale satellite images.

The experimental land cover/use classification with the neural network classifier shows:

The number of hidden units and layers affects the accuracy significantly. Both over
-
structured and under
-
structured neural networks can occur.

The increase of the learning rate reduces the learning time, but degrades the overall
accuracy, especially secondary succession in the study area.

The size of training samples does not affect significantly the accuracy of the
classification.

The classification accuracies using a neural network classifier are better overall than
using a maximum likelihood classifier. The neural network is especially superior when
few training samples are available.

Neural networks work well with multiple scale satellite images.

IKONOS

Landsat TM

ASTER

IKONOS (Left: true color; Right: false color)

The

study

area

is

located

between

W
52
º
31
'
5
",

S
2
º
59
'
25
"

and

W
52
º
3
'
3
",

S
3
º
30
'
16
"
.

This

is

a

typical

tropical

area

in

the

east

Amazon,

Brazil,

where

a

collage

of

cultivated

land,

pasture,

forest,

succession

features,

river,

and

city

exists
.

These

features

are

used

in

the

comparison

of

land

cover

classification
.


NN is more consistent in accuracy over
scales. NN did a better job than MLC
at all three scales in this context.

Genong (Eugene) Yu
a1
, Ryan R. Jensen
a2
, Paul W. Mausel
a3
, Eduardo S. Brondizio
b4
, Emilio F. Moran
b5
, and Vijay O. Lulla
a6
,

a.
Department of Geography, Geology and Anthropology, Indiana State University, Terre Haute, IN 47807, USA; b. ACT/Department of

An
thropology, Indiana University, Bloomington, IN 47405, USA.

(Emails: 1. g
-
yu@indstate.edu; 2. r
-
jensen@indstate.edu; 3. pmausel@scifac.indstate.edu; 4. Ebrondiz@indiana.edu; 5. Moran@indi
ana.edu; 6. Vijay_lulla@mama.indstate.edu.)


References


Duda, R.O., Hart, P.E., and Stork, D.G., (2001),
Pattern Classification
. John Wiley &
Sons, Co., 654p.


Paola, J. and Schowengerdt, R. A., (1995). A review and analysis of backpropagation
neural networks for classfication of remotely
-
sensed multi
-
spectral imagery.
International
Journal of Remote Sensing
, 16: 3033
-
58.


Foody, G. M., and Arora, M. K. (1997), An evaluation of some factors affecting the
accuracy of classification by an artificial neural network.
International Journal of Remote
Sensing
, 18(4):799
-
810.


Bishop, C.M., (1995),
Neural Networks for Pattern Recognition
. Oxford: Clarendon Press;
New York: Oxford University Press, 1995. 482p.