Title: Lecture 21 Remote Sensing III Obtaining Information from Satellite Imagery
1Lecture 21Remote Sensing III Obtaining
Information from Satellite Imagery
2Lecture Plan
- 1. Spectral image enhancement
- 2. Spatial image enhancement
- 3. Unsupervised classification
- 4. Supervised classification
3- What is Image Enhancement?
- Altering the appearance of an image in such a way
that the information of an image is better
interpreted visually in terms of a particular
need. - It may involve single bands of data or multiple
bands of data. - Enhancement may be spectral or spatial.
4Why and when do we enhance Imagery?
- In its basic form, digital imagery simply a
series of brightness values - x, y location, z brightness value per band
- When?
- Brightness values are clustered within a narrow
range, - Particular combinations of spectral bands are
required to discriminate between different
features, - We want to emphasize certain features and are
less interested in other feature types. - We want to remove small speckles of differing
reflectance - We want to increase the resolution of an image.
- Many others
51. Spectral Enhancement
6Linear Contrast Stretch
- Raw data values in an image are explored in order
to find best range of values for display of
particular features and aid their interpretation.
- e.g. Bands 2, 3 4 of a Landsat TM image of
Lowland rainforest and sugar cane farms.
Unenhanced
7Linear Contrast Stretch (cont.)
- The same image after a simple linear image
contrast stretch.
8Histogram Equalisation
- Non-linear form of contrast enhancement.
- Uses the whole histogram in the contrast
enhancement. - Classes with low frequency have been amalgamated
- Classes with high frequency have been spaced out
more widely than originally
9Band Ratios
The contrast in land-cover types in 1 or 2 can
be used to enhance image data. A transformation
involves
- BVs in a band being changed by applying a
formula
- or by combining BVs from more than one band and
applying a formula.
- NIR/ VISIBLE - a division or ratio.
- It enhances the differences between vegetated and
un-vegetated areas.
10Reasons for using Band Ratios
- Logically, ratioing will cancel out or reduce
whatever is common in two images and exaggerate
where they are different.
- This creates a new set of data that may be used
to highlight certain features.
11A ratio transformation does a number of things
- it enhances imagery
- it reduces 2 bands to one REDUNDANCY
- it can actually be used as a basis for physical
measurements - (crop productivity, LAI). Field measurements
correlated with ratio values. TONNES/ HA
- Thus, because of this stability they can be
used for monitoring
- Other ratios include TM3/TM2 as a measure of iron
oxide concentration and TM5/TM7 as a clay
alteration guide.
12- A band ratio is a new channel of data created by
the division of two sets of band digital numbers.
- It is the simplest of the multi-spectral
techniques, and a type of GIS 'overlay' - New ratio channel a (band x / band y ) (where
'a' is a numerical scaling factor)
13Spectral Slope Enhancement
- In general, band ratioing can emphasise the
difference between adjacent spectrum sections in
a single image - e.g. the reduction of the 'topographic effect'
- the Infra-red and red is the most common ratio.
- Since healthy vegetation involves high
reflectance in IR and low in red, any IR/Red (or
any visible wavelength) will show vegetation
differences more clearly
- Higher values (IR/red) more vegetation
(biomass)
14(No Transcript)
15Which other ratios would be useful?
- Generally pairs of bands that are highly
correlated will not show much information in
their ratios
- There are many options in a multi-band dataset
n (n-1) / 2 3 (3 bands), 6 (4 bands), 15 (6
bands), 21 (7 bands)
- We might more often pick pairs of bands from
different portions of the EM spectrum - MSS IR x Vis
- TM Vis x IR x MIR
16- However there are some useful applications using
two bands in the same region, e.g. in geology - MSS 5/4, 6/5 7/6 (4 green, 5 red, 6,7 NIR)
- 3/2, 7/5, 3/1, 5/7 mineral enhancement (hydro
thermally altered rocks)
- Ratio of two bands in same EM region distinguish
- soil types, geologic types
- Ratio of two bands not in the same area of the EM
spectrum distinguishes... - major groups e.g. water, rock, bare, coniferous,
deciduous
17Vegetation indices
- An index involves a 'normalised difference',
which compensates for additive effects as well as
multiplicative effects negated by ratioing.
- Band difference / band sum NDI or Normalised
Difference Index.
- The most common is the Normalised Difference
Vegetation Index (NDVI)
- NDVI (IR - Red) / (IR Red) For TM (TM4
-TM3) / (TM4 TM3)
- NDVI is used extensively to present a measure of
vegetation amount or biomass, especially in
regional and global estimates.
- Other 'normalised indices' include
- NDSI (TM2-TM5) / (TM2TM5) (S Snow)
- NDGI (TM4-TM2) / (TM4TM2) (G Green)
18Indices (cont.)
- The use of ratios as a monitoring function is
even better if they are NORMALISED. - Normal ratio scaling (0-255 / 0-255) is usually
between -5 and 5 and this range often differs
between dates.
- Normalised ratios have the same scaling between
dates -1 to 1 they are bounded ratios and can
thus be compared.
- This compatibility between dates enhances its
utility for comparing physical measurements
between dates.
- NDVI is important as a vegetation discriminator
19(No Transcript)
20NDVI
- It is also assumed to normalise reflectance from
the soil background in an area of vegetation and
also reflectance from the atmosphere so that all
that is being recorded is differences in
vegetation.
- It is not just a primary resource it also
represents an energy flux. Vegetation comprises
99 of earth's biomass. It is an important source
of global energy and moisture for atmosphere.
- Total surface area of leaf material in contact
with atmosphere is greater than surface area of
planet. Major solar energy exchange surface
through process of evapotranspiration (ET).
21- Composites of Ratios
- Ratio 1/4 for the blue band, basalt stands out
- Ratio 4/2 for green as this ratio depicts
vegetation - Band 3 for red as it was the only band not yet
used. - Basalts show up as blue
- Basaltic andesite shows up as pink
- Vegetation is green
22b. Empirically-based Image Transforms
- Specific transformations that are developed from
real data using specific sensors. - Originally based on Landsat MSS data in the
1970s, when researchers saw that agricultural
crops occupied an identifiable region of the 4
dimensional space of MSS bands. - ie. certain features (eg. soils, agricultural
crops) can be clearly identified upon specific
manipulation of the data. - 2 examples
- Perpendicular vegetation index
- Tasselled cap transform.
23Perpendicular Vegetation Index
- Plot of visible red vs. NIR for a partly
vegetated area - PVI72 (0.355MSS7 - 0.149MSS5)2 (0.355MSS5 -
0.852MSS7)2 - Value is correlated with green leaf area index
and biomass.
- Requires care, however.
- Small sample size
- Specific location
- Sensitive to recent rainfall
Conceptual diagram of the PVI, which is
proportional to the perpendicular distance from P
to the soil line S1-S2.
24Tasselled Cap (Kauth-Thomas) Transform
- Related to PVI
- Kauth and Thomas (1976) defined a linear
transformation to enhance the data according to
the data structure in a particular dataset. - Better represents soil and vegetation gradients
in multispectral data
Output Layer1 Brightness Layer2
Greenness Layer3 Wetness
25Principal Components Analysis
PCA output (3 Axes) Axis2 (green gun) relates to
natural vegetation structure
Landsat ETM (7 bands)
262. Spatial Enhancement
- Methods for selectively empasising or supressing
information at different spatial scales, or
artificially increasing the resolution of an
image. - For example,
- Suppress noise caused by detector imbalance
(banding) - Emphasise edges between fairly homogenous
features - Remove scattered bright pixels from a forest
canopy - Removal of noise such as scatter in a radar
image.
27Types of Spatial Enhancement
- 1. Filtering
- a. Low-pass (smoothing) filters.
- a. High-pass (sharpening) filters.
- 2. Edge detection
- 3. Frequency-domain filter
- 4. Resolution merge
- Can get to these from
- viewer/raster/filtering/convolution
- Interpreter/spatial enhancement
281a. Low-pass (smoothing) Filters
1 dimensional example Cross-section of a TM image
- Often, the underlying trends in data are obscured
by local patterns and random noise. - Therefore, removal.
- Involves a moving window of 3x3, 5x5 or 7x7
pixels, so that every pixel (except those at the
edge of the image) are allocated new values based
of those of their neighbours - Some uses
- Removal of banding.
- Smoothing away effects of image-to-image
mis-registration
29- This is an example of a moving average low-pass
filter. - The target pixel therefore becomes the average of
all those around it. - Another types include
- Median filters
- Adaptive filters based on the mean and variance
of the grey levels beneath the window. Eg. Sigma
filter
301a. Low-pass (smoothing) filters (cont).
Original image
55 low pass
33 low pass
31a. High-pass (sharpening) filters.
- These techniques de-blur an image
- Two Main Approaches
- Image subtraction
- Derivative-based methods
- Derivative-based methods use rate of change of
grey-scale value over space - These filters exaggerate difference where
gradients in change in BVs of adjacent pixels are
stronger.
322. Edge detection
- Edges are sharp changes in the grey-scale values
of an image or layer. - May have some interpretation regarding cultural
features such as roads. - Also very important in geology identifying
geological structure such as a fault. - A range of techniques
- 2 sets of filter matrices (x direction and y
direction) - Subtract low-pass filtered image from the
original - Others
332. Edge detection (cont)
Y weight matrix
X weight matrix
344. Resolution Merge
- Resolution Merge functions integrate imagery of
different spatial resolutions (pixel size). These - Can be used either intra-sensor (i.e., Landsat TM
multi with Landsat TM Pan) or inter-sensor (i.e.,
SPOT panchromatic with Landsat TM). - A key element of these multi-sensor integration
techniques is that they retain the thematic
information of the multiband raster image. - Thus, you could merge a Landsat TM (28.5 m pixel)
scene with a SPOT panchromatic scene (10 m pixel)
and still do a meaningful classification, band
ratio image, etc.
35Resolution Merge
Landsat ETM multi (30m) Landsat ETM Pan (15m)
Landsat ETM multi (15m)
36Image classification
- Digital image classification is process of
assigning pixels to classes according to some
classification scheme - Compare unknown pixels to one another and to
pixels of known identity or class using image
bands as variables - Classes form regions on image or map, so end
result is a mosaic of uniform patches - Process uses classification algorithms or
classifiers
37Image classification
- Classification may be end point (ie land
classification), - May be an intermediate step in the analysis (ie
further spatial and statistical analysis later) - Many different classification strategies
- Characteristics of each image and study differ,
therefore different classification strategies
need to be used for different tasks - For this reason, essential that the analyst
understands the different image classification
techniques to allow the most appropriate
classifier to be chosen.
38Types of Classification
- Unsupervised Classification
- Minimal interaction from the analyst.
- Searching for natural groups of pixels in the
image. - Supervised Classification
- Considerable interaction by the analyst.
- Identifies areas on an image that are known to
belong to each category. - Uses these areas to guide the classification.
39One band classification - density slicing
- Dividing a range of brightnesses into intervals,
then assigning each interval a colour. - Can be achieved through unsupervised or
supervised techniques.
NDVI Values
Split into classes
40Unsupervised Classification
- Identification of natural groups within
multispectral data. - Classification program
- automatically searches for natural groupings or
clusters of the spectral properties of pixels
41Unsupervised procedure
- Typical procedure for performing an unsupervised
classification is as follows - Run the program to cluster the image data into a
user-defined number of spectral classes - Display the output classified image and assign
each spectral class a tentative name and colour
corresponding to a thematic class - Pool and reclassify where multiple classes
represent an homogenous feature of interest - Evaluate the accuracy of the classification by
overlaying classes or using statistical methods
42Clustering generalisation
Fine (41 classes)
Broad (11 classes)
43Supervised Classification
- Using samples of known identity to classify
pixels of unknown identity into one of a number
of information classes. - These samples are known as training samples, and
are located in training areas. - Training samples can come from
- Ground surveys,
- Identification on the image
- Overlay or comparison with other spatial data
- Selection of appropriate training areas are vital
44Overall Project Plan Using Supervised Methodology
- State nature of problem
- Identify region of interest
- Identify classes of interest
- Address spatial, spectral and temporal resolution
issues - Acquire remote sensing and ground reference data
- Geometric rectification
- Issues of atmosphere, season, etc.
- a priori knowledge of area (surveys)
45Landsat TM example
- Supervised classification based on known
vegetation types - provides details of eleven structural vegetation
classes - can provide association mapping for floristically
simple situations - accuracy assessment gained from ground surveys
and API
46(No Transcript)