Lab for Remote Sensing Hydrology and Spatial Modeling Dept. of Bioenvironmental Systems Engineering, NTU Satellite Remote Sensing for Land-Use/Land-Cover.

台灣大學生物環境系統工程學系 Lab for Remote Sensing Hydrology and Spatial Modeling遙測水文及空間模式研究室 Dept. of Bioenvironmental Systems Engineering, NTU

衛星遙測在變遷偵測之應用Satellite Remote Sensing for

Land-Use/Land-CoverChange Detection

鄭克聲台灣大學生物環境系統工程學系


Change Detection

Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times.

In general, change detection involves the application of multi-temporal datasets to quantitatively analyze the temporal effects of the phenomenon.


Because of the advantages of repetitive data acquisition, its synoptic view, and digital format suitable for computer processing, remotely sensed data have become the major data sources for different change detection applications during the past decades.


Applications of Remote Sensing Change Detection

Land-use/land-cover changesForest or vegetation changeForest mortality, defoliation and damage assessmentDeforestation, regeneration and selective loggingWetland changeForest fireLandscape changeUrban changeEnvironmental changeOther applications such as crop monitoring, changes

in glacier mass, etc.


Factors Affecting Accuracy ofChange Detection

precise geometric registration between multi-temporal images

calibration or normalization between multi-temporal images

availability of quality ground truth datathe complexity of landscape and environments of

the study areachange detection methods or algorithms usedclassification and change detection schemesanalyst’s skills and experienceknowledge and familiarity of the study areatime and cost restrictions.


Lambin and Strahler (1994) listed five categories of causes that influenced land-cover change: long-term natural changes in climate

conditionsgeomorphological and ecological processes

such as soil erosion and vegetation succession

human-induced alterations of vegetation cover and landscapes such as deforestation and land degradation

inter-annual climate variabilitythe greenhouse effect caused by human

activities.


When selecting remote sensing data for change detection applications, it is important to use the same sensor, same radiometric and spatial resolution data with anniversary or very near anniversary acquisition dates in order to eliminate the effects of external sources such as sun angle, seasonal and phenological differences.


Before implementing change detection analysis, the following conditions must be satisfied: precise registration of multi-temporal

imagesprecise radiometric and atmospheric

calibration or normalization between multi-temporal images

similar phenological states between multi-temporal images

selection of the same spatial and spectral resolution images if possible.


Major Methods of Change Detection

Post-classification methodsImage-differencing methodsPrincipal component analysis methods


Statistical Perspective of Change Detection

Uncertainties involvedA statistical-test perspective

Null and alternative hypothesesTest statisticLevel of significance


Sources of Uncertainties inRemote Sensing Change Detection

Spatial and/or temporal variations in atmospheric conditionssoil moisture conditionsvegetation growth conditionsorographic conditions


The Soil and Water Conservation Bureau (SWCB) implements a standard process to routinely monitor land-cover changes on slopeland.

The process basically calculates grey level difference between two images and adopts a threshold value of grey level difference for land-cover change detection.

Image differencing on single band or composite images is the most widely used approach of change detection.


Problems and challenges How should the threshold value be

determined?How much confidence do we have on decision

of change detection?


Determining Threshold for Change DetectionMultiple of standard deviation of DN difference.

Nelson (1983): k = 0.5~1

Ridd and Liu (1998): k = 0.9~1.4

Sohl (1999): k = 2


Thresholding of grey level difference is globally based. It does not consider the grey level correlation of multi-temporal images and grey level of the pixel under investigation.

It is important to examine the bivariate scatter plot of multi-temporal images.


Bivariate Scatter Plot of Multi-temporal Images

Red band

01/10/1999 vs 21/09/2002






Pre- and post-images of the same spectral band are highly correlated.

Bivariate scatter plot shows bivariate joint probability distribution.


Change Detection Using Bivariate

Probability Contours95% probability contour

X2

X1

: detected changes


Conditional Prob. Distribution

Bivariate Joint Probability Distribution and Conditional

Probability Distribution X2

X1

Joint Prob. Distribution

)|( 12| 112xXf xXX


Class-specific Temporal Correlation

X2

X1


Transforming Change Detection to Hypothesis Test

Using conditional probability distribution, the work of change detection can be placed in the framework of hypothesis test.

Null hypothesis Ho: no change.


Bivariate Normal Distribution

Conditional normal distribution

Parameters can be estimated using pixels associated with no change.

Critical regions with respect to chosen level of significance can then be determined.


Major features (vegetation, soil and water) classification for individual images.

Class-specific correlation analysis using pixel pairs that are not associated with change.

Determining bivariate probability distribution for each class.

Specifying class-specific critical regions for test at level of significance .


01/10/1999

IR R G

water

vegetation

soil


21/09/2001

IR R G

water

vegetation

Soil


SPOT Image of the Study Area

Sept. 21, 2001Oct. 1, 1999


Confidence level 50%, 75%, 90%

Water

VegetationSoil


Detected Changes (R Band)

90% confidence region 95% confidence region


Detected Changes (IR/G)


Examples of Detected Changes

21/09/2001Changed sites 01/10/1999


Summary

We have demonstrated that change detection can be placed in a hypothesis test framework.

Preliminary results are promising.Several problems remain:

Non-gaussian probability distributionsUncertainty in parameters estimationDifficulty in deriving conditional

probability distributions.


Thanks for your attention.

Lab for Remote Sensing Hydrology and Spatial Modeling Dept. of Bioenvironmental Systems Engineering, NTU Satellite Remote Sensing for Land-Use/Land-Cover.

Documents

remote sensing hydrology

spatial modeling

cover change detection

change detection analysis

remote sensing data

decision of change detection

approach of change detection

landcover change detection