page 1 of 50

Optimization of Artificial Neural Networks in Remote Sensing Data

Analysis

Tiegeng RenDept. of Natural Resource Science in URI

(401) 874-9035

tren8835@postoffice.uri.edu

10/07/2002

page 2 of 50

Outline

• Introduction of Satellite Remote Sensing• Remote Sensing Image Classification

Methodology• Data and Experiments• Discussion and Conclusion

page 3 of 50

Earth Surface and Land Cover Map

How to achieve an accurate land cover map in natural resources research?

page 4 of 50

Remote SensingRemote Sensing is defined as: The science of acquiring, processing and interpreting images that record the interaction between electromagnetic energy and matter.

page 5 of 50

Different Type of Land Use and Land cover and Spectral Characteristics

Blue Band

Red Band Near-infrared Band

Green Band

page 6 of 50

From Digital Image to Land-use Map

Visualization Classification

• Multi-spectral Digital Image

• Pseudo Color Image

• Land-use and Land-cover Map

page 7 of 50

Classification System

In this study, we used the USGS classification with 10 land cover categories:

• 1. Turf/Grass

• 2. Barren land

• 3. Conifer forest

• 4. Deciduous forest

• 5. Mixed forest

• 6. Brush land

• 7. Urban land

• 8. Water

• 9. Non-forest wetland

• 10. Forest wetland

page 8 of 50

Remote Sensing Image Classification

• Statistical Methods

– Supervised– Unsupervised

• Artificial Neural Network Approaches• Rule-based• ...

page 9 of 50

Supervised Remote Sensing Image Classification

Vegetation

Water

Soil

Band 1 (0 ~ 255)

Band 2 (0 ~ 255)

page 10 of 50

Classification Process

MappingRelationship

Methods:

Statistical classifier

ANN-based classifier

Observation space

SolutionspaceLandsat TM

Band1Band2Band3Band4Band5Band6Band7

0~255 0~255 0~255 0~255 0~255 0~255 0~255

Category 1Category ..Category …Category …Category …Category …Category N

WaterwetlandForestAgri.UrbanResidential

4045611938011225

Category: Forest(Pattern)

page 11 of 50

Statistical Methods

- Need Gaussian (Normal) distribution on the input data which is required by Bayesian classifier.

- Restrictions about the format of input data.

Band 1 (0 ~ 255)

Band 2 (0 ~ 255)

Vegetation

Water

Soil

page 12 of 50

Statistical Methods:

Deciduous forest

Conifer forest

Forest Wetland

Mixed forest

Brush Land

Non-forest Wetland

How to find a boundary for the following patterns?

page 13 of 50

Artificial Neural Network Approach

• No need for normal distribution on input data• Flexibility on input data format• Improved classification accuracy• Robust and reliability

page 14 of 50

……………

…………

Artificial Neural Network Is Defined by ...

• Processing elements• Organized topological structure• Learning rules

page 15 of 50

Processing Element (PE)

f(x) OutputInput

PE

Artificial counterparts of neurons in a brain

Wj1

Wj2

Wj4

Wj3

Wj5

page 16 of 50

PE’s Output

Function of Processing Elements

……………

…………

unitj

o1

o2

o3

fwj1

wj2

wj3

oj

PE’s Inputs

• Receive outputs from each PEs locate in previous layer.

• Compute the output with a Sigmoid activation function F(Sumof(Oi*Wji))

• Transfer the output to all the PEs in next layer

page 17 of 50

Input layer Hidden layer Output layer

……………

…………

Input vector i(x1, x2, … xn)

Output vector i(o1, o2, … om)

Artificial Neural Network Architecture - Backpropagation ANN (BPANN)

Feed Information Analysis Result

page 18 of 50

Pattern Recognition

Vegetation:

(10, 89) ------> (1,0,0)

(11,70) ------> (1,0,0)

… … … (1,0,0)

Water:

(10, 21) ------> (0,1,0)

(15, 32) ------> (0,1,0)

… … … (0,1,0)

Soil:

(50, 40) -------> (0,0,1)

(52, 40) -------> (0,0,1)

… … … (0,0,1)

Band 1 (0 ~ 255)

Band 2 (0 ~ 255)Vegetation

Water

Soil

page 19 of 50

ANN Design

Pattern

Input layer Hidden layer Output layer

…………

How many PEs we need - Basic rules in designing an ANN.

• Input layer PEs - by dimension of input vector

• Output layer PEs - by total number of patterns (classes)

…………

Feed Forward

Back-Propagate

(0 ~ 255)(0 ~ 255)

(0 ~ 1)

(0 ~ 1)(0 ~ 1)

Band 1 (0 ~ 255)

Band 2 (0 ~ 255)Vegetation

Water

Soil

page 20 of 50

ANN Training - From Pattern to Land Cover Category

Pattern

…………

…………

Feed Forward

Back-Propagate

Vegetation:

(10, 89) ------> (1,0,0)

(11,70) ------> (1,0,0)

… … … (1,0,0)

Water:

(10, 21) ------> (0,1,0)

(15, 32) ------> (0,1,0)

… … … (0,1,0)

Soil:

(50, 40) -------> (0,0,1)

(52, 40) -------> (0,0,1)

… … … (0,0,1)

Land Cover Category

10

89

1 (Vegetation)

0 (Water)

0 (Soil)

A vegetation pixel(10, 89) ------> (1,0,0)

page 21 of 50

After Training

water

Vegetation

Soil

…………

…………

x

y

1 (Vegetation)

0 (Water)

0 (Soil)

A new pixel (x,y), x in band 1, y in band 2

A Well-trained ANN

page 22 of 50

Problems with Traditional Back-Propagation ANN Approaches

• Time-consuming– Always over 10000 iteration, over 5 hours for a small case.

• Black box - uncontrollable training– Easily trapped in local minimum.

• Training result unpredictable

page 23 of 50

Optimization Techniques to ANN Approach

– Data Representation - Transform to Binary and Gray code format.

– Improved Convergence Algorithms - Apply Conjugate Gradient and Resilient Propagation.

– Weight initialization - Linear Regression.

page 24 of 50

Data Representation

Data representation - Use Binary format and gray code format instead of integer value. • Avoid to compute in the saturation range of the activation function• Increased the computation space (from 6 to 48)• Make the value more smoothly - Gray Code

page 25 of 50

Saturation Range of Activation Function - Sigmoid Function

Saturation

Original Inputs:

0 ~ 255

New Format:

0 or 1

page 26 of 50

Coding Method

4045611938011225

Integer format Binary format

0 0 1 0 1 0 0 0 0 0 1 0 1 1 0 10 0 1 1 1 1 0 11 1 0 0 0 0 0 10 1 0 1 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 1 0 0 1

Forest Pixel

6 Integers 48 Integers

page 27 of 50

Gray Code vs. Binary Format

• Gray Code is another coding system, similar to binary

• Can better represent continuous value

Binary format Gray code formatInteger

127 0 1 1 1 1 1 1 1 0 1 0 0 1 0 0 0

128 1 0 0 0 0 0 0 0 0 1 0 1 1 0 0 0

page 28 of 50

Improved Convergence algorithms

To Make the convergence more robust and reliable, apply:

• Conjugate Gradient (CG) • Resilient Propagation (RPROP)

page 29 of 50

Error Space and Weight Adaptive

Total Error

Steps

Wij

Wkl

page 30 of 50

Weight Adaptive with Conjugate Gradient

Wij

Errorw

Global minimum

page 31 of 50

Resilient Propagation (RPROP)

page 32 of 50

Weight initialization

Cut convergence time, do weight initialization using linear regression to pre-process the internal weight, make ANN adapted to the target pattern before training.

page 33 of 50

Data and Experiments

• Study Area

• Data Distribution

• Fine tuning the ANN structure

• Classification Result

page 34 of 50

Study Area - Rhode Island 1999 Landsat-7 Enhanced Thematic

Mapper Plus (ETM+) Image

Band 4,3,2In RGB

Band 5,4,3In RGB

page 35 of 50

Training and Testing Pattern

Training sample and Testing sample

Class Name Training SampleSize (pixels)

Testing Sample Size(pixels)

Agriculture 116 153Barren Land 146 125

Conifer Forest 173 146Deciduous Forest 343 217

Mixed Forest 265 155Brush Land 75 75Urban Area 287 108

Water 238 133Non-forest Wetland 248 163

Forest Wetland 52 57Total Pixels 1943 1332

page 36 of 50

Distribution of Training Data

Patterns Plotted by band 3 x 4

Patterns Plotted by band 4 x 5

Deciduous forest

Turf / Grass

Barren land

Conifer forest

Forest Wetland

Mixed forest

Brush Land

Non-Forest Wetland

Water

Urban area

page 37 of 50

ANN Design and Tuning

• Number of PEs in input layer = Number of spectral Bands in the remote sensing image• Number of PEs in output layer = Number of land cover categories• Number of PEs in hidden layer(s) to be determined for best performance

page 38 of 50

ANN Design and Tuning- Before And After Optimization

page 39 of 50

Classification Result

page 40 of 50

Classification Result- A Close Look

Rhode Island 1999 ETM+ Rhode Island 1999 Land-use and Land-cover map

page 41 of 50

Accuracy Assessment

page 42 of 50

Discussion and Conclusion

• Accuracy

• Gray code Vs. Integer format

• Robust and reliability

• Discuss on number of hidden layer PEs

page 43 of 50

Accuracy Comparison

Non-optimized ANN Accuracy - 79.58%

Optimized ANN Accuracy - 92.27%

page 44 of 50

Gray-code vs. Integer Value

Comparison Between Gray Code and Integer Format

page 45 of 50

Resilient Propagation and Weight Initialization

Performance Comparison on Different Scenario

Algorithm Total Runs Out Runs Time/iterationPure BP ANN

(Conjugate gradients)30 6 >10000

Optimized ANN(Regression and

RPROP)

30 30 Average < 500

Improvement in robust and reliability

page 46 of 50

Hidden Layer PE Number’s Influence On Classifier’s Performance.

• 48-150-10 ( 48 inputs, 150 hidden neurons, 10 output classes)• 48-250-10 ( 48 inputs, 250 hidden neurons, 10 output classes)• 48-350-10 ( 48 inputs, 350 hidden neurons, 10 output classes)

page 47 of 50

Hidden Layer PE Number’s Influence On Classifier’s Performance.

350 hidden layer PEs provide the best performance

Benchmark on Number of Hidden Neurons

0%

20%

40%

60%

80%

100%

120%

Agricultu

re

Barr

en

Conife

r

Decid

uous

Mix

ed

Bru

sh

Urb

an

Wate

r

Non-f

ore

st W

etla

nd

Fore

st W

etla

nd

350 Hidden neurons

250 HiddenNeurons

150 Hidden neurons

page 48 of 50

Conclusion

• Coding method, especially the Gray code is important to Remote Sensing image classification.

• Optimized ANN provide a robust and reliable classification solution

• Optimized ANN lead to higher classification accuracy

• Improved natural resources mapping

page 49 of 50

Acknowledgement

The research was funded by NASA (Grant No.

NAG5-8829) to Dr. Y.Q. Wang

• Dr. Y.Q. Wang (major professor)• Dr. Pete August (NRS, committee member)• Dr. Ken Yang (EE&CE)• Dr. Tom Boving (GEO)• Lab for Terrestrial Remote Sensing• Environmental Data Center

page 50 of 50

Thank You!