1of 29 - PowerPoint PPT Presentation

1 / 42
About This Presentation
Title:

1of 29

Description:

1of 29 – PowerPoint PPT presentation

Number of Views:31
Avg rating:3.0/5.0
Slides: 43
Provided by: ren89
Category:
Tags: 1of | idoc

less

Transcript and Presenter's Notes

Title: 1of 29


1
Artificial Neural Network Application in Remote
Sensing
  • Tim Ren
  • M.S. Candidate
  • Dept. of Natural Resource Science in URI
  • (401) 874-9035
  • tren8835_at_postoffice.uri.edu
  • 03/09/2001

2
Outline
  • Overview of the Study
  • Artificial Neural Network(ANN) Introduction
  • Rhode Island 1999 Land-use and Land-cover
  • Classification
  • Web-Based ANN System

3
Part I Overview
Part I Overview

4
Remote Sensing Data Collection
  • Satellite Multispectral Data of EMR
  • Analog/Digital Transformation
  • Multi-band Digital Image (False color image)

5
Multi-source and Multi-spectral Spatial Data

Band1 Band2 Band N GIS Aerial
photo ...

Multisource spatial data provide information from
different perspectives in data modeling and
information extraction.
6
Part II Introduction to Artificial Neural
Networks (ANN)
  • Neuron or PE (perceptrons)
  • Neural network structures
  • Working mechanism of Back-propagation ANN (BPANN)
  • Practical optimization algorithms on BPANN

7
ANN Architecture ---
Processing Elements


PE Output
PE Input
unitj
wj1
o1
wj2
o2
oj
S, f
o3
wj3
8
ANN Prototypes and Working Mechanism
9
Back-propagation Mechanism
  • Compute total error
  • Compute the partial derivatives
  • Update the weights and go to next epoch
  • W (t1) W(t)

E

10
Back-propagation Mechanism

Error
Global minimum
Wij
11
Algorithms and Optimization Methods--- Update
the Weights by What Rule?
  • Normal Back-propagation
  • Conjugate Gradient Method
  • Delta-Bar-Delta Rule
  • SuperSAB
  • Quickprop
  • Rprop (Resilient Back-propagation)

12
Coding --- Effective Way to Represent Data
Integer format
Binary format
forest
0 0 1 0 1 0 0 0 0 0 1 0 1 1 0 1 0 0 1 1 1 1 0
1 1 1 0 0 0 0 0 1 0 1 0 1 0 0 0 0 0 1 1 1 0 0 0
0 0 0 0 1 1 0 0 1
40 45 61 193 80 112 25
  • Increase number of weights
  • Enlarge the computation space
  • Increasing calculate time

13
Coding --- Network Structure for Coded Dataset
Original Structure 6 input neuron
After coded 48 input neuron
14
Part III Real World Application --- RI 1999 ANN
Applied Classification
  • The original data
  • Process of ANN classification
  • Classification accuracy

15
Study Area --- Subset of RI 1999 ETM image
Band 4,3,2 In RGB
Band 5,4,3 In RGB
Rhode Island 1999 ETM
16
Flow of Two Different Classification Methods
Normal classification by ERDAS
ANN classification
Export to ASC
Neural Network training
NN engine process image (in ASC format)
Land-use Land-cover map (in ASC format)
Import to Pixel
17
How The Data Looks Like ?
Different classes plotting using band 3 x 4
Different classes plotting using band 4 x 5
18
How The Data Looks Like ?
Signature files of different classes plotting
using band 3 x 4 x 5
19
Classification Result
Rhode Island 1999 Land-cover and Land-use map
Rhode Island 1999 ETM
20
Classification Result Zoom In --- Kingston
Campus
Turf / Grass
Barren land
Conifer forest
Deciduous forest
Mixed forest
Brush Land
Urban area
Water
Wetland 1
Wetland 2
Rhode Island 1999 ETM
Rhode Island 1999 Land-cover and Land-use map
21
Accuracy Assessment
22
Part IV Web-Based Internet Implementation
  • Why we use Internet
  • How it works

23
System Architecture
24
1st layer --- Web Interface
Tcl/tk and Tclet
Training set files
Neural Network Parameters
Algorithm and Message Window
25
Communicate Layer --- Communication Protocol and
Agent
Send Client side program (embedded in
Interface) e.g. send EOS 12000 ..
Computation and classification result returned
to the message window
Tclet gathering variables and files Send transfer
them to Server
EOS
Web Server
Server (stand alone program) Listen to port
12000
Neural Computation Engine
26
End --- Neural Computation Engine
EOS Host
Responses back to user interface
Server Listen to port 12000
Neural Computation Engine
New process forked Parameters and training set
transferred
e.g. neural lt configuration files .
27
What To Do Next ?
  • The internal mechanism of the weighs update.
  • The trace of weight update
  • Relation between error surface and the training
    data set
  • More algorithms and components to the neural
    engine
  • Methodology improvements (modularize ANN? )
  • Effectively handle multi-user conditions on
    distributed architecture of the Web-ANN

28
Summary
  • We applied BP ANN in the classification of Rhode
    Island 1999 ETM data, we got reliable result.
  • Feed-Forward ANN is a powerful algorithm used in
    the remote sensing image process, however, a lot
    optimization must applied.
  • To expand the use of ANN, we developed a
    Web-based ANN interface and system, which is
    useful for not only the remote sensing
    classification but also in other data mining and
    data analysis.

29
Acknowledgement
  • Lab for Terrestrial Remote Sensing
  • Environmental Data Center
  • Dr. Y. Q. Wang
  • Dr. Y. Wang
  • NASA Grant No. NAG58829

30
Index of Appendixes
  • I BP ANN --- Error space, weight adaptive
    technologies
  • II Coding examples
  • III Different signature for different classes
    plotting
  • IV What input format to choose --- gray code
    VS.
  • integer value
  • V Benchmark --- how many neurons in hidden
  • layer gives the best result
  • VI Communicate agent --- send and server
  • VII Multi-user solution

31
Appendix I BP ANN---Error Space, Weight Adaptive
Technologies
Steps
Total Error
Wkl
Wij
32
Appendix II Coding Examples
In integer format
forest
( 40 45 61 193 80 112 25 )
Binary format
Gray code format
Integer
7 0 0 0 0 0 1 1 1 0 0 0 0 0 1 0 0 8
0 0 0 0 1 0 0 0 0 0 0 0 1 1 0 0 9 0 0
0 0 1 0 0 1 0 0 0 0 1 1 0 1
33
Coding Examples --- Classification
Process
Observation space
Solution space
Landsat TM Band1 Band2 Band3 Band4 Band5 Band6 B
and7
Mapping Relationship
0255 0255 0255 0255 0255 0255 0255
Category 1 Category .. Category Category
Category Category Category N
Water wetland Forest Agri. Urban Residential
Methods Linear Non-linear Statistical ANN
34
Coding Examples
  • Integer pattern

40 45 61 193 80 112 25
(Pattern Mapping)
Forest
  • Coded pattern

0 0 1 0 1 0 0 0 0 0 1 0 1 1 0 1 0 0 1 1 1 1 0
1 1 1 0 0 0 0 0 1 0 1 0 1 0 0 0 0 0 1 1 1 0 0 0
0 0 0 0 1 1 0 0 1
(Pattern Mapping)
Forest
35
Appendix III Different Signature for Different
Classes Plotting
All 6 Vegetation
Band 3 x 4
Band 4 x 5
36
Appendix III Different Signature for Different
Classes Plotting
Urban and Barren land
Band 3 x 4
Band 4 x 5
Barren Land
Urban area
37
Appendix IV What Input Format to Choose ---
Gray-code VS. Integer Value
Benchmark on different input format
Integer Format - 6 input neurons
Gray Code Format - 48 input neurons
38
Appendix IV What Input Format to Choose ---
Gray-code VS. Integer Value
39
Appendix V Benchmark --- How Many Neurons in
Hidden Layer Gives the Best Result
48-150-9 ( 48 inputs, 150 hidden neurons, 9
output classes)
48-250-9 ( 48 inputs, 250 hidden neurons, 9
output classes)
48-350-9 ( 48 inputs, 350 hidden neurons, 9
output classes)
40
Appendix V Benchmark --- How Many Neurons in
Hidden Layer Gives the Best Result
41
Appendix VI Communicate Agent --- send and
server
Socket
4
send eos 12000 add 1 3
  • Commands
  • 1.compute add, sub
  • 2.system commands start "Unix commands"
  • 3.stop command stop pid
  • 4.system load uptime,loadaver
  • 5.collect data fetch filename
  • 6.send file post filename
  • Example
  • send eos 12000 "add 1 2"
  • send eos 12000 start neural lt configure_file "

42
Appendix VII Multi-user Solution
Socket
Socket
1. Assign a new Work ID. 2. Create a unique
working directory to avoid file conflict. 3.
Clear temporary working directory
Write a Comment
User Comments (0)
About PowerShow.com