Tim Ren, M.S. Candidate

1

Artificial Neural Network Application in Remote
Sensing
Tim Ren, M.S. Candidate Department of Natural
Resources Science University of Rhode
Island 04/28/2000

2
Preview

• Introduction to Remote Sensing
• Objectives
• Why ANN (Artificial Neural Network)
• How ANN works
• ANN Application in Remote Sensing

3
Part I Introduction to Remote Sensing

4
Remote Sensing Data Collection

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

5
Remote Sensing of the Earth Surface
Improve classification performance How to
achieve an accurate land cover map?
6
Multi-source Spatial Data

Band1 Band2 Band N GIS Aerial
photo ...

Multisource spatial data provide information from
different perspectives in data modeling and
information extraction.
7
Digital Image Processing

Multi-source data
visualize
A picture
Classification result
classify
8

• Traditional method of classification
• Assumption
• Methodology
• - Unsupervised Classification
• - Supervised Classification
• Accuracy of Classification
• Limits

9

• Objectives
• Develop artificial neural network algorithms to
  handle multispectral and multitemporal remote
  sensing and multisource spatial data
• Build efficient ANN architecture, establish
  learning rules to train and refine ANN paradigms
• Apply the trained ANNs in remote sensing data
  modeling (classification and change detection)

10
Why ANN ?

• No need for the Gaussian (Normal) distribution
  about the input data (as required by Bayesian
  classifier)
• No need for the prior knowledge about the input
  data before the classification process
• No restrictions about the format of input data
  (More flexible and robust in multi-source spatial
  data classification A Promising alternative to
  Bayesian classification)

11
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
40 45 61 193 80 112 25
(Pattern)
Forest
12
Questions to Answer

- Does ANN algorithm perform better than
traditional statistical method? - Which ANN
paradigm is better (Backpropagation?
Modularized ANN?...) - How effective an ANN can
do in multisource spatial data analysis and
modeling?
13
Part II Introduction to Artificial Neural
Networks



14
Artificial Neural Network Is Defined by ...

• Processing elements
• Organized topological structure
• Learning rules

15
Artificial Neural Network Is Defined by ...

• Processing elements
• Organized topological structure
• Learning rules

16
Processing elements (PE)

Artificial counterparts of neurons in a brain
PE
Wj1
Wj2
output path
Wj3
?
f(x)
input
Wj4
Wj5
17
ANN Architecture -Processing Elements


PE Output
PE Input
unitj
wj1
o1
wj2
o2
oj
S, f
o3
wj3
18
Artificial Neural Network Is Defined by ...

• Processing elements
• Organized topological structure
• Learning rules

19
ANN Architecture -Topological Structure
Input layer
Hidden layer
Output layer


Input vector i(x1, x2, xn)
Output vector i(o1, o2, om)
20
Organized topological structure


21
Organized topological structure
--Back-Propagation ANN Architecture
Land-cover Categories
Output layer Hidden layer Input layer
Landsat TM, GIS...
22
Artificial Neural Network Is Defined by ...

• Processing elements
• Organized topological structure
• Learning rules

23
ANN Architecture - Learning Rules
Input layer
Hidden layer
Output layer


Input vector i(x1, x1, xn)
Output vector i(o1, o1, on)
How the ANN learns ?
24
Supervised Learning with a Teacher
- Paired training set ( Input and Output)
25
Unsupervised Learning


- Self-Organize
26
Reinforcement Learning


- Learning with Critic
27
Part III ANN application in Remote Sensing



?
28
ANN application in Remote Sensing

- Multi-source Spatial Data Classification -
Change Detection - Land Cover Change and
Prediction
29
ANN applied in the Remote Sensing

- Multi-source Spatial Data Classification
Remote Sensed Data
Classification Result land cover map
30
- Multi-source Spatial Data Classification

Input layer
Hidden layer
Output layer


Remote Sensed Data
Grassland
Woodland

Wetland

Other source GIS, Airphoto .

Urban
31
ANN Applied in Remote Sensing
- Change Detection
Changes between 1985 - 1997
1985
1997
32
- Change Detection
(2mn1o) network
Chang Map with Complete Land Cover
Change Information
Land cover change extractor
Image A
Image B
33
ANN Application in Remote Sensing
- Land Cover Change and Prediction
1980
1990
2010
?
34
Plan of Research

• - Study Area
• - Data
• Landsat TM
• GIS
• Field Observation
• (USGS EROS Data Center)
• - Design of Artificial Neural Network

35
Summary
Data from Other sources
Remote Sensing Data
36
Acknowledgement
Dr. Y.Q.Wang Dr. Yong Wang NASA Grant No.
NAG58829 Apr. 28 2000

37
Artificial Neural Network Is Defined by ...
Input layer
Hidden layer
Output layer


Input vector i(x1, x1, xn)
Output vector i(o1, o1, on)
unitj
wj1
o1
Units in next layer
wj2
o2
oj
S, f
o3
wj3
38
Artificial Neural Network Is Defined by ...
Input layer
Hidden layer
Output layer


Input vector i(x1, x1, xn)
Output vector i(o1, o1, on)
unitj
wj1
o1
Units in next layer
wj2
o2
oj
S, f
o3
wj3
39
Artificial Neural Network Is Defined by ...
Input layer
Hidden layer
Output layer


Input vector i(x1, x1, xn)
Output vector i(o1, o1, on)
unitj
wj1
o1
Units in next layer
wj2
o2
oj
S, f
o3
wj3
40
Artificial Neural Network Architecture


41
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
Methods Linear No-linear Statistical ANN