Transcript Unsupervised Classification

UNIVERSIDAD DE GUADALAJARA
CENTRO UNIVERSITARIO DE LOS VALLES
Archaeological Land Use Characterization
using Multispectral Remote Sensing Data
Dr. Iván Esteban Villalón Turrubiates, Member, IEEE
María de Jesús Llovera Torres
Monitoring Hidrological Variations using Multispectral
SPOT-5 Data: Regional Case of Jalisco in Mexico
Dr. Iván Esteban Villalón Turrubiates, Member, IEEE
Overview
- Abstract
- Remote Sensing Definition
- Sensor Resolution
- Introduction to Image Classification
- Model Formalism
- Verification Protocols
- Simulation Experiments
- Concluding Remarks
Abstract
 Proposition - A new and efficient classification approach of
remote sensing signatures extracted from large-scale
multispectral imagery.
 Contribution - This approach exploits the idea of
combining the spectral signatures from a remote sensing
image to perform a novel and accurate classification
technique.
 Verification - Simulation results are provided to verify the
efficiency of the proposed approach.
REMOTE SENSING
DEFINITION
Remote Sensing
 Remote Sensing can be defined as:
 "The arte and science to obtain data from an object
avoiding direct contact with it” (Jensen 2000).
 There is a transmission medium involved?
Remote Sensing
 Of the Environment:
 … is the collection of information regarding our Planet
surface and its phenomena involving sensors that are
not in direct contact with the studied area.
The main focus is in recollected information from a spatial
perspective throughout electromagnetic radiation
transmission.
Remote Sensing
 Sensor election.
 Reception, storage and digital signal
processing of the data.
 Analysis of the resulting information.
Process
A) Illumination Source
B) Radiation
C) Interaction with the object
D) Radiation sensing
E) Transmission, reception and data processing
F) Analysis and interpretation
G) Application
SENSOR RESOLUTION
Resolution
 All remote sensing systems use four types of
resolution:
 Spatial
 Spectral
 Temporal
 Radiometric
Spatial Resolution
Spectral Resolution
Temporal Resolution
July 2
July 8
August 3
16 days
Time
11 days
July 1
July 12
July 23
August 3
Radiometric Resolution
6-bits Range
0
63
8-bits Range
0
255
10-bits Range
0
1023
INTRODUCTION TO IMAGE
CLASSIFICATION
Image Classification
 Why classify?
 Make sense of a landscape
 Place landscape into categories (classes)

Forest, Agriculture, Water, Soil, etc.
 Classification scheme = structure of classes
 Depends on needs of users.
Typical uses
 Provide context
 Landscape planning or assessment
 Research projects
 Natural resources management
 Archaeological studies
 Drive models
 Meteorology
 Biodiversity
 Water distribution
 Land use
Example: Near Mary’s Peak
•Derived from a 1988 Landsat TM image
•Distinguish types of forest
Legend
Open
Semi-open
Broadleaf
Mixed
Young Conifer
Mature Conifer
Old Conifer
Classification: Critical Point
 LAND COVER not necessarily equivalent to LAND USE
 We focus on what’s there: LAND COVER
 Many users are interested in how what is there is being
used: LAND USE
 Example
 Grass is land cover; pasture and recreational parks are
land uses of grass
Basic Strategy: How to do it?
 Use radiometric properties of remote sensor
 Different objects have different spectral signatures
Basic Strategy: How to do it?
 In an easy world, all “vegetation” pixels would have
exactly the same spectral signature.
 Then we could just say that any pixel in an image
with that signature was vegetation.
 We could do the same for soil, water, etc. to end up
with a map of classes.
Basic Strategy: How to do it?
But in reality, that is not the case. Looking at several pixels with
vegetation, you’d see variety in spectral signatures.
The same would happen for other types of pixels, as well.
The Classification Trick:
Deal with variability
•Different ways of dealing with the variability lead to
different ways of classifying images.
•To talk about this, we need to look at spectral signatures a
little differently.
Think of a pixel’s brightness in a
2-Dimensional space. The pixel
occupies a point in that space.
The vegetation pixel and the
soil pixel occupy different
points in a 2-D space.
With variability, the
vegetation pixels now
occupy a region, not a point,
of n-Dimensional space.
Soil pixels occupy a
different region of
n-Dimensional space.
Basic Strategy:
Deal with variability
• Classification:
• Delineate boundaries of classes in n-dimensional space
• Assign class names to pixels using those boundaries
Classification Strategies
 Two basic strategies:
 Supervised Classification
 We impose our perceptions on the spectral data.
 Unsupervised Classification
 Spectral data imposes constraints on our interpretation.
Supervised Classification
Supervised classification requires the analyst to
select training areas where he knows what is
on the ground and then digitize a polygon
within that area…
The computer then creates...
Mean Spectral
Signatures
Conifer
Known Conifer
Area
Water
Known Water
Area
Deciduous
Known Deciduous
Area
Digital Image
Supervised Classification
Mean Spectral
Signatures
Information
Multispectral Image
(Classified Image)
Conifer
Deciduous
Water
Unknown
Spectral Signature of
Next Pixel to be
Classified
The Result: Image Signatures
Land Cover Map
Legend:
Water
Conifer
Deciduous
Unsupervised Classification
 In unsupervised classification, the spectral data imposes
constraints on our interpretation.
 How? Rather than defining training sets and carving out
pieces of n-Dimensional space, we define no classes
beforehand and instead use statistical approaches to
divide the n-Dimensional space into clusters with the
best separation.
 After the fact, we assign class names to those clusters.
Unsupervised Classification
The analyst requests the computer to examine the
image and extract a number of spectrally distinct
clusters…
Digital Image
Spectrally Distinct Clusters
Cluster 3
Cluster 6
Cluster 5
Cluster 2
Cluster 1
Cluster 4
Unsupervised Classification
Saved Clusters
Cluster 3
Output Classified Image
Cluster 6
Cluster 5
Cluster 2
Cluster 1
Cluster 4
Next Pixel to
be Classified
Unknown
Unsupervised Classification
The result is essentially the
same as that of the
supervised classification:
It is a simple process to regroup
(recode) the clusters into
meaningful information classes
(the legend).
Land Cover Map
Legend
Labels
Water
Water
Water
Conif.
Conifer
Conifer
Hardwood
Hardwood
Hardw.
MODEL FORMALISM
Multispectral Imaging
 Is a technology originally developed for space-based imaging.
 Multispectral images are the main type of images acquired by remote
sensing radiometers.
 Usually, remote sensing systems have from 3 to 7 radiometers; each
one acquires one digital image in a small band of visible spectra,
ranging 450 to 690 nm, called red-green-blue (RGB) regions:
 Blue -> 450-520 nm.
 Green -> 520-600 nm.
 Red -> 600-690 nm.
 The combination of the RGB spectral bands generates the so-called
True-Color RS images.
Weighted Pixel Statistics Method
 Statistical Approach.
 Assume normal distributions of pixels within classes.
 For each class, build a discriminant function
 For each pixel in the image, this function calculates the
probability that the pixel is a member of that class.
 Takes into account mean and variance of training set.
 Each pixel is assigned to the class for which it has the
highest probability of membership.
Weighted Pixel Statistics Method
Mean Signature 1
Candidate Pixel
Mean Signature 2
It appears that the candidate pixel is
closest to Signature 1. However, when we
consider the variance around the
signatures…
Blue
Green
Red
Near-IR
Mid-IR
Weighted Pixel Statistics Method
Mean Signature 1
Candidate Pixel
Mean Signature 2
The candidate pixel clearly belongs to the
signature 2 group.
Blue
Green
Red
Near-IR
Mid-IR
Weighted Pixel Statistics Method
Weighted Pixel Statistics Method
VERIFICATION PROTOCOLS
Verification Protocols
 A set of three synthesized images are used as verification
protocols.
 All synthesized images are True-Color (RGB), presented in
1024-by-1024 pixels (TIFF format).
 Each synthesized image contains three different regions (in
yellow, blue and black colors) with a different pattern.
 The developed Weighted Pixel Statistics (WPS) algorithm
is compared with the most traditional Weighted Order
Statistics (WOS) method [S.W. Perry, H.S. Wong, 2002].
Results:
1st Synthesized Scene
Synthesized Scene
WOS Classification
WPS Classification
Quantitative Comparison
1st Synthesized Scene
Results:
2nd Synthesized Scene
Synthesized Scene
WOS Classification
WPS Classification
Qualitative Comparison
2nd Synthesized Scene
Synthesized Scene
WOS Classification
WPS Classification
Quantitative Comparison
2nd Synthesized Scene
Results:
3rd Synthesized Scene
Synthesized Scene
WOS Classification
WPS Classification
Qualitative Comparison
3rd Synthesized Scene
Synthesized Scene
WOS Classification
WPS Classification
Quantitative Comparison
3rd Synthesized Scene
Remarks
 The quantitative study is performed calculating the classified
percentage obtained with the WOS and WPS methods,
respectively.
 The WOS method uses only 1 spectral band.
 The WPS method uses the information from the three spectral
bands to analyze the pixel-level neighborhood means and
variances.
 The results shows a more accurate and less smoothed
identification of the classes.
SIMULATION EXPERIMENTS
Archaeological Land Use
 A Remote Sensing Signatures (RSS) electronic map is extracted
from the multispectral image. Three level RSS are selected for this
particular simulation process, defined as:
 ██ – Archaeological land use zones.
 ██ – Modern land use zones.
 ██ – Natural land cover zones.
 ██ – Unclassified zones.
Archaeological Site
"Guachimontones", Jalisco Mexico
Simulation Results
Scene from "Guachimontones"
Original Scene
WPS Classification
Hidrological Variations
 A Remote Sensing Signatures (RSS) electronic map is extracted
from the multispectral image. Three level RSS are selected for this
particular simulation process, defined as:
 ██ – Humid zones.
 ██ – Dry zones.
 ██ – Wet zones.
 ██ – Unclassified zones.
Simulation Results
Scene from "La Vega" dam, Jalisco Mexico
Original Scene
WPS Classification
CONCLUDING REMARKS
Remarks
 The WOS classifier generates several unclassified
zones because it uses only one spectral band in
the classification process.
 The WPS classifier provides a high-accurate
classification without unclassified zones because it
uses more robust information in the processing.
 The qualitative and quantitative analysis probe the
efficiency of the proposed approach.
Future Work
 Comparison with several classification techniques.
 A more extensive performance analysis of the
proposed approach with different synthesized
images.
 Application to remote sensing imagery and the
study of its performance.
 Hardware
approach.
implementation
of
the
proposed
UNIVERSIDAD DE GUADALAJARA
CENTRO UNIVERSITARIO DE LOS VALLES
THANK YOU!
Questions?
Dr. Iván Esteban Villalón Turrubiates, Member, IEEE