change detection and time series analysis - geo … · change detection and time series analysis...
TRANSCRIPT
Vor
lesu
ng A
llgem
eine
Fer
nerk
undu
ng, P
rof.
Dr.
C. S
chm
ulliu
s
Geoinformatik & Fernerkundung, Friedrich-Schiller-U niversität Jena
Change detection and time series analysis
Lecture by Martin HeroldWageningen University
Overview
1. Types of land dynamics
2. Methods for analyzing multi-temporal remote sensing data:
• Change detection
• Time series analysis
3. Selected examples (land/forest cover)
4. Coarse vs. fine scale satellite data for monitoring land dynamics
Land dynamics and
Change Detection
Why assess landscape change with remote sensing imagery?
– natural resource inventory/monitoring
– explain environmental phenomena
– develop/validate explanatory (process) models
– support integrated landuse/landcover change analysis and accounting applications
http://www.tiem.utk.edu/bioed/webmodules/circadianrhythm.html
Daily
Alan Belward Ispra 22 Feb,, 22 Apr., 22 Jul.22 Oct. 2005
Seasonally
http://www.ronet.com.br/marrocos/pv-antig/pv1-18.html
http://www.skyscrapercity.com/showthread.php?t=344422
Porto Velho, Brazil 2007Porto Velho, Brazil 1908
Annually
Shanghai, China 1987 Shanghai, China 2004
Michael Glantz NCAR
Types of change:
• short term change (synoptic weather events)
• cyclic change (seasonal phenology)
• directional change (urban development)
• multidirectional change (deforestation & regeneration)
• event change (catastrophic fires, disasters)
Change Detection
The accuracies of remote sensing based change detection depend on:
1.Spectral, spatial, and temporal characteristics of the multi-date
imagery,
2.Precise registration, calibration, or normalization between multi-
temporal images,
3.Complexity of investigated landscape and the related processes
affecting land change dynamics,
4.Change detection methods or schemes used,
5.Analyst’s experience and familiarity of the study area, and
availability of quality ground reference data, e.g. for accuracy
assessment,
6.Time and cost restrictions.
Change detection algorithms 1Change detection
algorithm
Approach Comments
Manual change
interpretation
Two images (or photographs)
are visually compared to
manually identify changes.
� This method is simple and sometimes the most feasible
method for change detection, especially when other techniques
fail to accurately identify changes
� Requires human experience � Results are subjective and strongly depend on interpretation
skills
� Time consuming interpretation processPost-classification
comparison
Independent spectral
classification results from each
image date and pixel-by-pixel
or segment-by-segment
comparison to detect changes
in land cover types.
� Minimizes radiometric differences between multi-date images
� Provides complete change matrix� The accuracy of the post-classification comparison is
dependent on the accuracy of the initial classification.
� Requires time and classification expertise
Composite
analysis
Multi-date images are analyzed
through joint classification
including change categories.
� Simple and time-saving in classification� Requires many classes (5 land cover classes=25 possible
change classes). � Demands prior knowledge of the logical interrelationships of
classesUnivariate image
algebra
(difference/ratio)
Subtraction/ratio of one date
original or transformed
imagery (e.g. vegetation index,
radiance, etc.) from a second
date image.
� Can only differentiate change or no-change, i.e. cannot
generate detailed change matrix information� Requires criterion/threshold of changes or no-changes (e.g.
Standard deviation method)
� Requires strict radiometric corrections
Change detection algorithms 2Change detection
algorithm
Approach Comments
Bi-temporal linear
data transformation
Use image transformation (e.g. Tasseled Cap, PCA) of multi-date composite image to identify changes (e.g. in brightness, greenness or wetness).
� Reduces data redundancy and emphasized differences
between images� Results might be hard to interpret� Requires threshold of change magnitude between change
and no change� Cannot generate detailed change matrix information
Change vector
analysis
Multivariate change detection technique that possesses the full dimensionality (spectral + temporal) of the image data and produces two outputs: change magnitude and change direction. A change has occurred if the vector surpasses a specific threshold value. The type of change that has occurred can be determined by the angle or direction of the change vector.
� Can produce detailed change detection information. Strict requirement for reliable image radiometry.
� Requires threshold of change magnitude between change and no change
� Difficult to identify change trajectories
Image Regression A mathematic model that describes the fit between two multi-date images. The dimension of the residuals is an indicator
of where change occurs.
� Can only differentiate change or no-change, i.e. cannot generate detailed change matrix information. Requires only initial radiometric image calibration/inter-
calibration
Multi-temporal
spectral mixture
analysis
Using spectral unmixing to detect changes in fraction images.
� Image fractions have biophysical meaning� Results are stable accurate and repeatable� Advanced image processing skills required� Requires threshold of change magnitude between change
and no change
Temporal Image Differencing (Band Subtraction)
The temporal image differencing procedure subtracts the DNs from one image date from another.
The differences in areas of little or no change will be small while areas of extreme change will be represented by large positive or negative changes in DNs (-255 / +255).
Temporal Image Differencing (Band Subtraction)
Pros:
• Band subtraction is simple to process
• The technique can be easily replicate across sites and time periods
• Continuous change values are created by the differencing
Temporal Image Differencing
Deviations from an index: (Image Differencing)
From Jensen http://www.cla.sc.edu/geog/rslab/rsccnew/rscc-frames.html
Temporal Image Differencing
Deviations from an index: (Image Differencing)
From Jensen http://www.cla.sc.edu/geog/rslab/rsccnew/rscc-frames.html
Forest cover 1973
Forest cover 1985
Change Vector Analysis (CVA)
Change vector analysis (CVA) is a multidimensional extension of the image differencing technique.
The difference between radiometric values (DNs) for multiple bands are calculated as change differences exhibiting vectors (change direction) and magnitudes (change intensity).
Change Vector Analysis (CVA)
One case (a) exceeds a change threshold is indicated with a vector arrow. The second case (b) depicts a pair of pixels whose radiometric values where determined to be within a threshold range and would be classified as no (insignificant) change.
Change Vector Analysis (CVA)
When multiple bands are used in the analysis, secto rs of change (quadrants) can be identified to depict different type of change (++, +-, --, -+).
Bands in the change vector analysis may represent preprocessed indices or data transformations. For example; The Tasselled Cap transformation yields an initial image with bands related to Greenness, Wetness, Brightness and Haziness.
When this type of transformation is applied before the CVA analysis, then change sectors may for example represent increases or decreases in any combination of bands.
Change Vector Analysis (CVA)
Four possible sectors of change derived by examining 2 bands of data.
Change Vector Analysis (CVA)
An interesting potential value of CVA which we are now exploring is the possibility of identifying trajectories of change over time.
For example: hypothetical trajectory of change depicting forest clearing, initial regeneration, ca nopy closure, and species change (conifer to deciduous).
The potential for assessing the “signature” for different types of ecological processes over time adds an interesting dimension to this type of analysis.
Change Vector Analysis (CVA)
An hypothetical trajectory of change depicting forest clearing, initial regeneration, canopy closure, and species change (conifer to deciduous).
Time Series Analysis
• Variation in images across different times:
– Cyclical variation (aka “seasonal” variation)
• LAI in deciduous ecosystems
• Temperature through the year, or throughout the day
– Trend variation (aka “long term change in mean”)
• LAI in human impacted ecosystems across years
• Temperature across years
– Other variation: random or non-random changes
across time
Cyclical Variation
• All ecosystems show a
degree of cycling, related
to, among other things,
evapotranspiration, day
length and temperature
differences throughout the
year.
• We want to characterize
period, amplitude and
offset of the cycles.
Examples/Techniques
Duration of growing
season
Principal Components Analysis (PCA)
• Rotation and scaling along orthogonal
directions of maximum variance
• Large correlation among multi-temporal
datasets – first component:
– state (Principal components)
– dynamics (seasonality, longer term trends) in
higher components
PC1
PC2
Consider multitemporal NDVI (1986-1990, monthly data):
Expect high degree of correlation
but also deviations from this
use PCT...
Monthly NDVI - Africa
96.68% of variance in PC1
Loadings very similar for all months
…average
Monthly NDVI - Africa
2% of variance in PC2
Dec-March minus
April-Nov
Seasonality - ITCZ movement
Observing change using coarse resolution data
1. Synergy of existing map products + ancillary data
– Challenging but some regional examples
2. Vegetation continuous change (VCC)
– Tree canopy cover/deforestation (20x20 km blocks)
3. Observation of active fire and burned areas
– Several operational products
4. Observing long-term trends
– AVHRR/NDVI times series data since 1981
– Night-time lights
5. Near-real time observations (i.e. deforestation/DETER)
Usefulness:
– Indicators and hot spots of change
– Guide more detailed analysis and true area estimations
– Understand inter-annual versus intra-annual dynamics
– Independent global measurements
Hot spots of forest cover change 1980-2000
Lepers et al., (2005). A synthesis of information on rapid land-cover change for the period 1981-2000. BioScience, 55 (2), 115-124
Longer term observations of vegetation
Relationship between positive temperature anomalies (left) and the trends in vegetation growth (right) for different
seasons periods 1998-2005 in Northern Eurasia based on SPOT Vegetation data
Huettich et al., 2006 IJRS
Global active fire observations
• Animated figure!
http://modis-fire.umd.edu/MCD45A1.asp
Contact: Luigi Boschetti <[email protected]>
EXAMPLE APPLICATIONS
• 1 year of composite of MODIS
burned areas, superimposed on
surface reflectance to provide
geographic context.
• Burned area statistics for the
same period, for vegetation type
Africa
0.00E+00
1.00E+05
2.00E+05
3.00E+05
4.00E+05
5.00E+05
6.00E+05
Jul-0
1 BA
Aug-01 B
A
Sep-01 B
A
Oct-01 B
A
Nov-01 B
A
Dec-0
1 BA
Jan-0
2 BA
Feb-02
BA
Mar-02 B
A
Apr-0
2 BA
May-0
2 BA
Jun-0
2 BA
fire
affe
cted
are
a [k
m^2
]
0%
5%
10%
15%
20%
25%
30%
unm
appe
d [%
]
croplands
barren_or_sparsely_vegetated
grasslands
savannas
woody_savannas
open_shrublands
closed_shrublands
mixed_forests
deciduous_broadleaf_forest
deciduous_needleleaf_forest
evergreen_broadleaf_forest
evergreen_needleleaf_forest
other
unmapped BA
DMSP nighttime light density: 1992 / 1998 / 2003
Night-time light data 1992-2003
Pan-humid tropics forest clearing, 2000-2005
Courtesy of M.Hansen, SDSU
Annual Gross Deforestation (Brazil)
Sinop – Mato Grosso, Brazil.
Source: INPE (2006)
• Annual rate: 11-29,000 km 2/ano• Total gross deforestation: 681.343 km 2
Ann
ualR
ate
-A
rea
(km
2 /an
o)
21050
17770
13730
29059
18161
13227
19400
27200
2375023266
18226
17259
17383
1816513786
11030
14896
0
5000
10000
15000
20000
25000
30000
35000
77/88
*88
/8989
/9090
/9191
/9292
/94 **
94/95
95/96
96/97
97/98
98/99
99/00
00/01
"01/0
2""0
2/03"
"03/0
4""0
4/05"
Summary and outlook
Observing types of change:• short term change (synoptic weather events)• cyclic change (seasonal phenology)• directional change (urban development)• multidirectional change (deforestation & regeneration)• event change (catastrophic fires, disasters)
Different types of earth observation data useful ba sed on spatial, temporal and thematic detail they provi de
Only specific changes are (usually) important for l and assessment and accounting purposes
Earth observation data is key source for studying spatio-temporal land processes (GIS, models)
Deforestation patterns and processes