change detection
DESCRIPTION
Change detection. Paul Aplin School of Geography, University of Nottingham, UK Chairman, Remote Sensing and Photogrammetry Society. Outline. Remote sensing background Multitemporal analysis Preprocessing requirements Change detection Image differencing Further reading - PowerPoint PPT PresentationTRANSCRIPT
Change detection
Paul Aplin
School of Geography, University of Nottingham, UK
Chairman, Remote Sensing and Photogrammetry Society
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Outline
Remote sensing background Multitemporal analysis Preprocessing requirements Change detection Image differencing Further reading Introduction to practical exercise
There are meaningful distinctions between remote sensing ‘platforms’, ‘sensors’ and ‘images’
Platform The craft on which a sensing
device is mounted
Sensor The sensing device or instrument
itself
Image The image data acquired by the
sensing device
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Platforms, sensors and images
There are three main categories of remote sensing platforms
Spaceborne- Satellite- Shuttle Ground-based
- Hand-held- Raised platform
Airborne- Aeroplane- Helicopter- Hot air balloon- Air ship- Tethered balloon
Commonestplatforms
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Remote sensing platforms
Satellite path
Field of view
Ground track (imaged area)
Advantages
Continuous data acquisition Permanent orbit
High geometric accuracy Stable orbit (no atmosphere)
Wide area of coverage High vantage point
Low data cost?
DisadvantagesGeometric distortion
Earth curvature
High operation cost Launch, etc.
Low spatial detail? High vantage point
Cloud cover?
High data cost?
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Satellite platforms
Generally, remote sensing satellites are in low Earth orbits (LEOs), at altitudes of several hundreds of kilometres
These satellites orbit the Earth approximately every hour
Most remote sensing satellites follow a ‘polar’ orbital path (approximately north-south)
Polar orbits maximise the area of data acquisition, exploiting the Earth’s rotation
Satellite orbit
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
As the Earth rotates eastwards, the satellite passes North-South (or South-North) acquiring a ‘swath’ of imagery
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Polar orbit
As the Earth rotates eastwards, the satellite passes North-South (or South-North) acquiring a ‘swath’ of imagery
As the Earth continues to rotate, another image swath is acquired
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Polar orbit
As the Earth rotates eastwards, the satellite passes North-South (or South-North) acquiring a ‘swath’ of imagery
As the Earth continues to rotate, another image swath is acquired
And again…
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Polar orbit
Some remote sensing satellites follow a ‘geostationary’ orbital path
This means they constantly view the ‘same’ area of coverage
By orbiting in coincidence with the Earth’s rotation
All geostationary satellites orbit around the Earth’s equator
A common example is meteorological satellites
Other non-remote sensing satellites also use geostationary orbits
E.g., communications
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Geostationary orbit
Advantages
High spatial detail Low vantage point
On-demand acquisition Requested flights
Low operation cost?
Avoid cloud cover?
Low data cost?
Disadvantages
Narrow area of coverage? Low vantage point
Sporadic acquisition Occasional flights
Low geometric accuracy Yaw, pitch, roll
High data cost?
Field of view
Ground track (imaged area)
Flight path
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Aeroplane platforms
There are various types of sensors or instruments
Any type of sensor can be operated from any remote sensing platform Satellite, aircraft or ground-based
Digital- Sensor- Camera- Video- Radar- LiDAR
Analogue- Camera
Described in lecture
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Sensors
The most common types of remote sensing instruments are the digital sensors introduced previously
Hyperspectral sensors are also fairly common
Remote sensing at these parts of the electromagnetic spectrum (visible, infrared) is collectively known as ‘optical’ remote sensing
Panchromatic sensorsCreate images comprising
one broad spectral waveband
Multispectral sensorsCreate images comprising
several spectral wavebands
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Digital sensors
Incoming solar radiation
Atmospheric distortion
Reflected radiationScattered
radiation
Received radiation
Sensor
Data download
User
Data supply
Sun
Earth’s surface
Ground receiving stationAbsorbed/transmitted
radiation
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Image acquisition
Radar = ‘Radio Detection And Ranging’
Radars are well known as navigational instruments, but also operate as remote sensing instruments
Using ‘microwave’ (long wavelength) part of electromagnetic spectrum
Radar remote sensing is very unlike optical remote sensing…
Radars are ‘active’, generating their own energy source Optical sensors are ‘passive’, using the sun’s reflected energy
Each radar operates at a specific spectral wavelength Optical sensors average reflectance across spectral wavebands
Common radar sensors Synthetic Aperture Radar (SAR) – generally spaceborne Side-Looking Airborne Radar (SLAR) – generally airborne
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Radar sensors
Radar instruments are ‘active’ and generate their own energy
Radars emit microwaves and record the ‘return’ from features
Radar images distinguish features on the basisof ‘texture’
Rough surfaces – high return (light image) Smooth surfaces – low return (dark image)
Radars have strong viewing capabilities Penetrate cloud due to long wavelength Operate at night due to own energy source
Emitted microwaves
Some return
Scattered microwaves
Radar
Little/ no return
Forest‘rough’ surfaceWater
‘smooth’ surface
Radar image acquisition
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Three main image characteristics are considered here
The term ‘resolution’ is only partly appropriate in some of these cases, but is used for convenience
Spectral resolution
Temporal resolution
Spatial resolution
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Remotely sensed images
Strictly speaking, spectral ‘resolution’ refers to the ‘width’ of a spectral waveband used to generate an image band
It is perhaps more useful to consider the ‘number’, ‘position’ and ‘width’ of all spectral wavebands comprising an image
• Spectral image characteristics have been covered in earlier description of electromagnetic spectrum and optical/radar imagery
Typical spectral reflectance curves
Band
1 -
blue
Band
2 -
gree
n
Band
3 re
d
Band
4 –
near
infra
red
Band
5 –
mid
infra
red
Band
7 –
mid
infra
red
Multispectral image
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Spectral resolution
In simple terms, spatial resolution means ‘pixel size’
The concept is simple…
Fine spatial resolution image = fine spatial detailsmall feature identification
Coarse spatial resolution image =coarse spatial detaillarge feature identification
As spatial resolution is degraded progressively, features become increasingly blurred and harder to identify
4 m
8 m
16 m
32 m
64 m
128 m
256 m
512 mSkukuza,Skukuza,Kruger National Park, Kruger National Park, South AfricaSouth Africa
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Spatial resolution
Spatial resolution is related to the area of coverage of an image (the ‘swath width’ of the sensor)
Both spatial resolution and swath width may be affected by the altitude of the platform…
High altitude (e.g., satellite) = wide swath = coarse spatial resolution
Low altitude (e.g., aircraft) = narrow swath = fine spatial resolution
However…
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Area of coverage
Swath width = 3 kmSpatial resolution = 500 m
No. pixels = 24Storage space = e.g., 6 Mb
Swath width = 6 kmSpatial resolution = 1 km
No. pixels = 24Storage space = e.g., 6 Mb
…altitude is not the only factor that determines spatial resolution and swath width
E.g., certain satellite sensors have fine spatial resolutions and narrow swath widths
Another problem is the limited processing and storage capacity of the instruments
Each image pixel requires a certain amount of storage space, regardless of spatial resolution
For instance, two images… with ‘different’ swath widths/spatial resolutions but the ‘same’ number or pixels (rows, columns) will require the ‘same’ amount of storage space
Therefore, there is a ‘play off’ between spatial resolution and swath width
Fine spatial resolution = narrow swath width Wide swath width = coarse spatial resolution
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Data processing and storage
In practical terms, swath widths and spatial resolutions can vary considerably
From images extending across 1,000s kilometres with pixels measuring 10s kilometres…
…through images extending across 100s kilometres with pixels measuring 10s metres…
…to images extending across 100s metres (or less) with pixels measuring 10s centimetres (or less)
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Practical examples
• Temporal resolution refers to the frequency of image acquisition at a constant location Other, synonymous terms are ‘repeat cycle’ or ‘revisit time’ This concept is most applicable to satellite sensors, since airborne remote sensing occurs
on an irregular basis
A collection of images over time of a constant location are known as a ‘multitemporal’ image or data set
As with spatial resolution, the concept is simple… Short temporal resolutions enable frequent observation of dynamic features Long temporal resolutions enable occasional observation of long-term change
The chief determinant of temporal resolution is satellite orbit, but also significant are tilting capabilities, latitude and swath width
Certain sensors can ‘tilt’ away from their normal orbital ground track, observing targeted features relatively frequently
‘Polar’ orbits lead to frequent coverage at high latitudes (e.g., polar regions) Wide swath sensors cover large areas, returning relatively frequently to any location
Therefore, by extension, there is also a ‘play off’ between temporal resolution and spatial resolution!
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Temporal resolution
In practical terms, temporal resolutions can vary considerably
From every half hour for meteorology sensors… …to weeks or even months for narrow swath sensors The key is to select suitable imagery related to the
feature to be monitored Hurricane FrancesOrbView-2 image
1 Sep 20041 Sep 2004
6 Sep 20046 Sep 2004
5 Sep 20045 Sep 2004 4 Sep 20044 Sep 20043 Sep 20043 Sep 2004 2 Sep 20042 Sep 2004
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Monitoring example
September 1972 December 1978 March 1987 June 1990
Multitemporal Landsat imagery
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Multitemporal imagery Multispectral image = image comprising multiple spectral wavebands
of a common area Multitemporal image data set = multiple images acquired at different
times of a common area
September 1972 December 1978 March 1987 June 1990
Multitemporal Landsat imagery
Normalized difference vegetation index
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Multitemporal analysis Any form of analysis involving multitemporal imagery
To enable multitemporal analysis, images acquired at different times must share certain common properties
Often image preprocessing is necessary Preprocessing refers to initial procedures conducted to prepare imagery
for subsequent analysis Main types of preprocessing:
Radiometric calibration Cosmetic operations (correcting for missing lines, etc.) Atmospheric correction Illumination correction Geometric correction (and/or registration) Terrain correction
For multitemporal analysis Geometric registration is essential to ensure that the images overlay each
other precisely Other corrections may be useful to ensure that any differences in the
images are not a result of instrument deterioration, atmospheric interference, variations in illumination (time of day/year), and geometric and topographic distortion
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Preprocessing
Atmospheric distortion affects remotely-sensed imagery, contributing (erroneously) to pixel values
Atmospheric correction is necessary where reflectance values are desired (as opposed to simple DNs) and where images are being compared over time
There are many methods for atmospheric correction, none of which is perfect and some of which are very complex
Relatively simple and common methods include: Dark object subtraction Histogram matching
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Atmospheric correction
The most common atmospheric effect on remotely-sensed imagery is an increase in DN values due to haze, etc. This increase represents error and should be removed Dark object subtraction simply involves subtracting the minimum DN value in the image from all pixel values
This approach assumes that the minimum value (i.e. the darkest object in the image) ‘should be’ zero The darkest object is typically water or shadow
Range (DN)
Freq
uen
cy
40(Min)
204 (Max)
67.12 (Mean) 0 DNs 255
Band 2
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Dark object subtraction
Histogram matching is a very simple and fairly crude method of atmospheric correction This involves adjusting one image to approximate another by aligning their histograms
This approach assumes that the ‘general’ differences between the images are due to external (e.g., atmospheric) effects Histogram matching nullifies these general differences, and remaining differences represent ‘real’ change between images
2001 1992Hist mtch 2001to1992
Band 1 Band 1Band 1
Min 42
Max 254
Mean 67
SD 10
Min 59
Max 254
Mean 74
SD 8
Min 59
Max 254
Mean 75
SD 8
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Histogram matching
Geometric registration involves registering an image without locational information to a known map coordinate system
This can be done by co-registering one image to another This often involves: 1. Identifying ground control points
2. Resampling the imageUnregistered image
1
2 3
N
Scale
12
3
Registered image
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Geometric registration
One common type of multitemporal analysis is change detection
Change detection involves the direct comparison of two or more images to identify how areas change over time
Various methods of change detection are possible…
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Change detection
• Image ratioing Division of one image by another Operates on individual image bands Areas of ‘change’ may be thresholded (e.g., > +25% and < -25%)
Post-classification comparison Each image is classified independently, reducing preprocessing need Resulting classifications are compared to identify change Change detection results affected by accuracy of input classifications
Change vector analysis Compares spectral properties of image pixels between dates Expresses change as vectors in spectral space
Composite multitemporal image analysis Images are combined to create a single multi-band image E.g., two 4-band images would create an 8-band composite image Various forms of analysis may be employed (classification, PCA)
Perhaps the most common and simple method, though, is…4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Change detection methods
Image differencing or image subtraction simply involves the subtraction of one image from another
No change = zero, change = +ve or –ve values
Operates on individual image bands
Areas of ‘change’ may be thresholded (e.g., > +25% and < -25%)
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Image differencing
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Reading for further information Aplin, P., 2006, On scales and dynamics in observing the environment,
International Journal of Remote Sensing, 27, 2123-2140. Camps-Valls, G., et al., 2008, Kernel-based framework for multitemporal
and multisource remote sensing data classification and change detection, IEEE Transactions on Geoscience and Remote Sensing, 46, 1822-1835.
Canty, M.J., 2007, Image Analysis, Classification and Change Detection in Remote Sensing, with algorithms for ENVI/IDL, CRC Press.
Chuvieco, E. (editor), 2008, Earth Observation of Global Change: The Role of Satellite Remote Sensing in Monitoring the Global Environment, Springer.
Coppin, P., et al., 2004, Digital change detection methods in ecosystem monitoring: a review, International Journal of Remote Sensing, 25, 1565-1596.
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Practical exercise Change detection using Idrisi image processing system
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Data Time series of NDVI imagery of Africa from the NOAA AVHRR sensor
Dec 87 Mar 88 Jun 88 Sep 88 Dec 88
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Image differencing
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Image thresholding
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Image regression
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Help – replacement files Backup files are provided
File to be created
Backup file provided
Description
DIFF8877 PADF8877 Difference image: AFDEC88 – AFDEC87CHG8877 PACH8877 Thresholded image: AFDEC88 – AFDEC87CHGMAP1 PACHMP1 Change map using thresholded image
SAMP87 PASAMP87 Contracted (degraded) image of AFDEC87
SAMP88 PASAMP88 Contracted (degraded) image of AFDEC88
TMP PATMP Step one (first scalar process) towards creating predicted map
PRED88 PAPRED88 Predicted map derived through regression
DIFFADJ PADFADJ Difference image: AFDEC88 – PRED88CHGADJ PACHADJ Thresholded image: AFDEC88 – PRED88CHGMAP2 PACHMP2 Change map using regression-derived thresholded image
Change detection
Paul Aplin
School of Geography, University of Nottingham, UK
Chairman, Remote Sensing and Photogrammetry Society
End of lecture
During the 2005 Ashes, the betfair Blimp showed aerial television coverage throughout the series
Old Trafford, Manchester11-15 August
Match drawn
But lest we forget…
England win series12 September
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
The betfair blimp
The earliest known examples of remote sensing involved taking photographs from (un-manned) balloons tethered above the area of interest
Boston 13 October 1860
Photographed by JamesWallace Black
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Historical remote sensing
For years, Meteosat images were downloaded by a ground receiving station in The University of Nottingham
The station is located at the top of the Tower building
The images were provided free to the meteorological/remote sensing community, and accounted for a large proportion of the university’s website traffic
After years of service, the station was deactivated in late 2003 due largely to lack of support and funding
Europe 11 August 1999
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
University of Nottingham METEOSAT ground receiving station
Aircraft movement can lead to severely geometrically distorted imagery In this case, aircraft ‘roll’ has caused sensor movement ‘across’ the
image swath, resulting in irregular edges
River Arun, West Sussex 22 July 1997
Compact Airborne Spectrographic Imager (CASI)
4th ISPRS Student Consortium and WG VI/5 Summer School, Warsaw 13-19 July 2009.
Aerial image distortion