Sie sind auf Seite 1von 47

Remote Sensing Classification Methods

Introduction to Remote Sensing


Example Applications and Principles

What is Classification? Explore and Classify an Image with MultiSpec

Questions

Definitions
Lillesand and Kiefer (1994) The science and art

of obtaining information acquired by a device that is not in contact with the object

CCRS Glossary A group of techniques for

collecting image or other forms of data from measurements made at a distance from the object, and the processing and analysis of the data.

Examples of Remotely Sensed Data


Weather Ocean Properties Physical Geography Major Disturbance Events / Hazards Cultural Features urban mapping Many other examples thematic info

back

disturbance

back

back

Decision

From Lillesand and Kiefer (1994)

O3 absorption at 0.2 um

H2O absorption at 1.4, 1.6, and 1.9 um

CO2 absorption at 2.0 um

CH4 absorption at 2.2 2.5 um

Human eye

E(l)i [ E(l)a + E(l)t ] = Er(l)


Incident Energy Reflected Energy

Images vs. Maps

What are the characteristics of a thematic map?

Abstraction Finite number of specified (discrete) classes Scale (coarse) Sharp boundaries cartographic tradition, vector representation

Finite number of specified classes


Generalize to a hierarchy of classes 1. Urban
A. Residential B. Commercial C. Industrial

2. Agricultural
A. Cropland B. Pasture C. Orchards

1. Etc

Scale
Scale of map determines the scale (and detail) of classes

100 Pixel Size (meters) 10 1 I II III

Level of Detail (Coarse to Fine)

Scale
Scale of imagery determines detail of what one sees
Forest Community types Individual tree species Branches, leaf City Neighborhood Block House (roof) & Garden

Sharp boundaries
Sharp
Political boundary Geological contact (unconformity)

Fuzzy
Vegetation community Wetland to dry area

In remote sensing: One class per pixel, and thus sharp boundaries

How to go from remote sensing data to classes?


Humans strong pattern recognition ability
Spatial
Shape Size Context

Spectral
Relative brightness Color

Remote Sensing
Usually purely spectral: each pixel classified independently

Snow Vs. Clouds

Snow Vs. Clouds


Clouds scatter at all wavelengths

Snow absorbs at >1.4 m


c

Remote Sensing Pixel Data


pixel 1,1 = B1 B2 B3 . . . z z z x x z z z z z x x x z z x x z x y y y y y y y y y z Band 1

Band 2

Bn

y = forest, x = agricultural field, z = natural grassland

Remote Sensing Pixel Data


Decision Boundary B1 B2 B3 . . . z z z x x z z z z z x x x z z x x z x y y y y y y y y y z Band 1 pixel 1,1 =

Band 2

Bn

y = forest, x = agricultural field, z = natural grassland

Classification Approaches
Supervised
Analyst identifies representative training sites for each informational class Algorithm generates decision boundaries

Unsupervised
Algorithm identifies clusters in data Analyst labels clusters

Supervised Classification
Analyst identifies training areas for each informational class Algorithm identifies signatures (mean, variance, covariance, etc) Classify all pixels Informational Class Map

Unsupervised Classification
Algorithm clusters data find inherent classes Classify all pixels based on clusters Spectral class map Analyst labels clusters (May involve grouping of clusters) Informational class map

Labeling of classes
Bispectral plots
Band 2

x x x x x

y y y y
Band 1

1 0.9 0.8

Reflectance (%)

Spectral curves

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6

Quartz Calcite Dolomite Gypsum Kaolinite

Wavelength (micrometers)

Classification map

Spectral Classes Vs. Informational Classes


Ideal Spectral classes A B Reality A B C Or A B C 1 2 1 2 3 4 Fresh blacktop Old blacktop Informational Classes 1 2 Concrete

Unsupervised Classification Methods


Histogram peak/valley finding ISODATA moving means

Histogram peak/valley finding


Histogram of values in the image
How many classes present?

Frequency

DN Value of Band x

Assumptions: More pure pixels than mixed pixels Classes are distinctive

Histogram peak/valley finding


Histogram of values in the image
How many classes present?

Frequency

DN Value of Band x

How do we identify significant peaks?

ISODATA
Iterative Self Organizing Data Analysis Technique More robust User specifies
Iteration
Maximum # Minimum change between iterations Maximum # (less usually found) Starting locations (along diagonal, or 1st PC, or arbitrary) Minimum # of pixels per cluster (e.g. 0.01% of the data, then merge) Maximum spread (e.g. if standard deviation large, then split)

Clusters

Splitting and merging parameters

ISODATA Classification: 1
Initial classification from arbitrary cluster centers Minimum distance to means classification rule
x

x x Band 2 x x x x x x x

Total distance = 77

+
Band 1

Arbitrary cluster centers

ISODATA Classification: 2
Initial classification from arbitrary cluster centers Minimum distance to means classification rule
x

x x Band 2 x x

+
x x x x x

Total distance = 54

Adjusted cluster centers

Band 1

ISODATA Classification: 3
Initial classification from arbitrary cluster centers Minimum distance to means classification rule Total distance = 26

x x Band 2 x x x x x x

+ x

+ x

Band 1

ISODATA Classification: 4
Initial classification from arbitrary cluster centers Minimum distance to means classification rule Total distance = 18

x x Band 2 x x x x x x

+ x

+x

Band 1

ISODATA Classification: 4
Initial classification from arbitrary cluster centers Minimum distance to means classification rule Total distance = 18 No changes in pixels between classes, therefore stop

x x Band 2 x x x x x x

+ x

+x

Band 1

MultiSpec RS Software
Research and education Long period development and refinement Freely available Supported by tutorials
http://cobweb.ecn.purdue.edu/~biehl/MultiSpec/

Das könnte Ihnen auch gefallen