USE OF HIGH RESOLUTION SATELLITE DATA

1
USE OF HIGH RESOLUTION SATELLITE DATA FOR CHANGE
DETECTION IN URBAN AREAS
F. Del Frate , G. Schiavon, C. Solimini
Università Tor Vergata - DISP Rome, Italy
2
NEURAL NETWORK APPROACH
Advantages

• Build internal decision rules directly from the
  data

• No need of a-priori statistical assumptions

• Possible effective sinergy between experimental
  data and data simulated by electromagnetic
  modelling

• Good properties of robustness and flexibility

multi-layer perceptron with a single hidden
layer and nonlinear activation functions is
capable of approximating any real-valued
continuous function, provided a sufficient number
of units within this hidden layer exists. K.
Hornik, M. Stinchcombe, A. White, Multilayer
feedforward networks are universal approximators,
Neural Networks, 1989
3
MLP Neural Network topology
So a nice tool but .
4
Handle with care !!!
The overfitting problem
5
Optimum training time
6
Optimum net topology
7
PROBLEM 1
Input
Landsat-TM
Output
3 Classes sealed, unsealed, water
Data set consisted of
35 Landsat images
50 Urban areas
26 Capital cities
8
Spectral analysis
Water surface
Band 1
Band 2
Band 3
Band 4
Band 5
Band 6
9
Spectral analysis
High Density residential
10
Algorithm implementation
Input pre-processing with normalization procedures
About 70000 spectral signatures used for training
(extracted from 14 images)
Optimized fast learning procedures (Scale
Conjugate Algorithm)
Final selected topology 6-9-9-3
Rate of processing for new images 700 pixels per
second
The same trained net is used for all new images
11
Berlin
12
New York
13
Rio de Janeiro
14
Tokyo
15
PROBLEM 2
Input
Quickbird Multispectral R, G, B, IR
Output
4 classes bare soil, asphalt, vegetation,
building
Test Area Tor Vergata University Campus and
surroundings
Data Set two images taken on 29 May 2002 and 9
March 2003
Same purpose as before to design one single net
to process both images
16
Acquisition Date 9/3/2003
17
29 May 2002
18
9 March 2003
19
(No Transcript)
20
Change detection
21
The NEURANUS Tool
The software Neuranus (NEURAl Networks
User-friendly Simulator) is a tool based on
neural networks for image processing

• Developed within an IDL environment
• Based on a window user interface
• User Friendly poor knowledge required about the
  neural networks theory
• Provides as a result a pixel-based
  classification of the selected image
• Produces the result in real time or near real
  time (depend of the inputs size)

The following four steps are included in the
software

• Generation of a statistically meaningful set of
  training data
• Definition of network topology
• Training phase
• Application of the trained net to the entire image

22
Principal dialogue window of Neuranus
input image
classified image
Control panel
Running the software example of the use of
NEURANUS
23
(No Transcript)
24
(No Transcript)
25
(No Transcript)
26
(No Transcript)
27
(No Transcript)
28
(No Transcript)
29
(No Transcript)
30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
New utilities !!