sec5 form r.ppt [read-only] - unisi.itgarzelli/documenti/07-sar-filtering-enhancement.pdf · canada...

Canada Centre for Remote Sensing, Natural Resources Canada

Radiometric Enhancement-Outline-

FilteringSpeckle Reduction

- Definition; Why speckle filtering; What is the ideal speckle reduction filter

- Non-adaptive filters (FFT filters)- Adaptive filters (Frost, Lee, MAP Gamma filters)

Edge Detection- Ratio edge detector filter- Touzi filter


Radiometric Enhancement (cont’d)-Outline-

Analysis of Image Texture

Visual Enhancement

Contrast Enhancement

Linear Enhancement

Nonlinear Enhancement

- Histogram, Exponential, Logarithmic, Power Law Stretch


IntroductionThis section reviews the methods of enhancing the radiometrics of an image using speckle reduction filters, spatial enhancement filters and visual enhancements.

The understanding of radar “speckle” is key to the understanding of SAR and SAR radiometric enhancements.

Often the reduction of speckle is desired to improve classification and/or for enhancement.

To reduce speckle, adaptive filters (e.g. map gamma filter), should be used rather than non-adaptive filters (e.g. FFT filters) on radar imagery.

Adaptive filters take into account the local properties of the terrain backscatter or the nature of the sensor, whereas non-adaptive filters do not.


Introduction to Speckle

Image variance or “speckle” is a granular noise that inherently exists in SAR imagery (Figure 5.1).

Speckle gives a single look image a grainy, salt and pepper appearance and is the dominating factor in radar imagery.

Speckle noise occupies a wider dynamic range than the scene content itself.

Images processed with a small number of 'looks' will have distribution intensities which are quite asymmetric due to speckle noise.

Creating a symmetrical histogram may not be the optimum procedure. Instead, pixels are set to the extreme limits of thedata intensity distribution (e.g. DN values of 0 and 255 for 8-bit data).

For a detailed review of speckle, see Raney (1998).


What is Speckle?

Speckle is coherent interference of waves scattered from terrainelements observed in each resolution cell.

An incident radar wave interacts with each element of the surface and surface cover to generate scattered waves propagating in all directions.

Those scattered waves that reach the receiving antenna are summed in direction and phase to make the received signal. The relative phase components contain the differential propagation paths.

The SAR focusing operation coherently combines the received signals to form the image.

The scattered wave phase addition results in both constructive and destructive interference of individual scattered returns and randomly modulates the strength of the signal in each resolution cell.


Figure 5.1 - Example of Speckle


What is Speckle? (cont’d)

Addition of backscatter from a collection of scatterers produces random constructive and destructive interference, see Figure 5.2.

Constructive interference is an increase from the mean intensity and produces bright pixels.

Destructive interference is a decrease from the mean intensity and produces dark pixels.

These random fluctuations give rise to speckle.

Reducing these effects enhances radiometric resolution at the expense of spatial resolution.


Figure 5.2 - SpeckleConstructive Interference

Destructive Interference

Result

Result

Example of Homogenous Target

Constructive interference

Destructive interference

Varying degrees of interference(between constructive and destructive )

Coherentradar waves


Speckle Suppression

Speckle results from a coherent (phase included) process.

Speckle can be reduced by incoherent (amplitude or power) processes.

Speckle reduction (or smoothing) necessarily reduces the resolution (increases the resolution cell size) of single channel SAR data.

Two basic linear processes:

- Multi-look - divides the signal into minimally overlapped frequency bands, processes each to a reduced resolution image, registers these, detects and adds the detected images. Examples of multi-look processing are shown in Figure 5.3.

- Averaging - detects the full resolution image, performs local averaging and resampling processes to create reduced resolution,reduced speckle images.

- For distributed targets both processes are equivalent.


Figure 5.3 - Multi-look ProcessingExamples of multi-look processing. Note that image chips A, B, and C all have the same resolution, but that image chips C and D have comparable image quality factors (data from an X-band airborne SAR, 1972, optically processed).(In Principles & Applications of Imaging Radar, Manual of Remote Sensing, 1998, Chapter 2 - Raney, pg. 75)Courtesy R.

Shuchman and E. Kasischke,

ERIM

A 6.1 m x 6.1 m N = 1

QSAR = 0.027

C 6.1 m x 6.1 m N = 16

QSAR = 0.43

B 6.1 m x 6.1 m N = 4

QSAR = 0.11

D 1.5 m x 2.13 m N = 1

QSAR = 0.31


Why Speckle Filtering?

The presence of speckle noise must be considered when selecting analysis methodologies.

Speckle filtering will permit:

better discrimination of scene targets.

easier automatic image segmentation.

the application of the classical enhancement tools developed for imagery from optical sensors such as; edge detectors, per-pixel and textural classifiers.


The Ideal Speckle Reduction Filter

Reduce speckle with minimum loss of information

In homogeneous areas, the filter should preserve:

radiometric information

edges between different areas

In textured areas, the filter should preserve:

radiometric information

spatial signal variability: textural information


Families of Speckle Reduction FiltersNon-adaptive filters

The parameters of the whole image signal are considered.Do not take into consideration the local properties of the terrain backscatter or the nature of the sensor. Not appropriate for filtering of non-stationary scene signal. Examples are the FFT filters.

Adaptive filtersAccommodate changes in local properties of the terrain backscatter.

- The speckle noise is modelled as being stationary - The target signal is not stationary since the mean backscatter

changes with the type of targetExamples are the Frost, Lee, Map Gamma, local mean and local median filters

Figure 5.4 shows examples of adaptive filters.


Figure 5.4 - Gamma vs. Median Filter

Tapajós, BrazilMay 20, 1996 Beam F2

Original Image

Median 5x5

Map Gamma 5x5


Kernel Size

Examples of Mean, Median and Mode filter kernels (or windows) are shown in Figure 5.5.

Filters are a sub-array of X by Y pixels that moves through the image.

All three filters shown in Figure 5.5 are square box filters, with a kernel size of 3 by 3 pixels

Degree of smoothing is a function of the size of the kernel.

As filter kernel size increases, smoothing increases.


Figure 5.5 - Filtering Kernel

Source: CCRS

5 7 49 8 65 5 8

MEAN

5 7 49 8 65 5 8

MEDIAN

5 7 49 8 65 5 8

MODE

5+7+4+9+8+6+5+5+8= 5757÷ 9 = MEAN = 6

4,5,5,5,6,7,8,8,9MEDIAN = 6

45556 MODE = 57889

3 x 35 x 5

7 x 7


Mean and Median FiltersPrinciple

Intensity at each sample interval in the image is replaced by the mean of pixel values in a moving window surrounding the sample.

The box or mean filter preserves well the radiometry but blurs textured areas.The median filter assigns the window median value to each sample.

Preserves texture information better Modifies the radiometric information of homogeneous areas, and does not preserve point target signature

Not recommended for radar imagery.See Figure 5.6 for examples of both filters.


Figure 5.6 - Median and Mean Filters


Original Image

Median 7x7

Mean 7x7


Adaptive Filtering

Adaptive filters (e.g. Map Gamma) reduce speckle while preserving the edges (sharp contrast variation).

Adaptive filters modify the image based on statistics extracted from the local environment of each pixel.

Larger kernel size (e.g. 11x11) result in an important increased smoothing effect on the resulting image (Figure 5.7).


Figure 5.7 - Gamma Filter


Original Image

Map Gamma 7x7

Map Gamma 11x11


Advantages of Adaptive Filters

Most of the well known adaptative filters require the calculation of the local observed mean and normalized standard deviation (coefficient of variation).

The adaptive filter produces an accurate estimate of the backscattering coefficient inside homogeneous (stationary) areas while preserving edge and texture structure in nonstationary scenes.


Most Well-known Filters: The Frost Filter

Principle

The unspeckled pixel value is estimated using a subwindow of the processing window.

The size of the subwindow varies as a function of target local heterogeneity measured with coefficient of variation:

– the larger the coefficient of variation, the narrower the processing subwindow

The Enhanced Frost Filter (Lopes, Touzi and Nezri, IEEE, 1990) minimizes the loss of radiometric and textural information (Figure 5.8).


Figure 5.8 - Examples of Filters


Most Well-known Filters : The Lee Filter

Principle

The unspeckled pixel value is a weighted sum of the observed (central) pixel value and the mean value.

The weighting coefficient is a function of local target heterogeneity measured with the coefficient of variation.

The Enhanced Lee Filter (Lopes, Touzi and Nezri, IEEE, 1990) minimizes the loss of radiometric and textural information (Figure 5.8).

The Enhanced Lee and Enhanced Frost Filters perform similarly.


Most Well-known Filters : The MAP Gamma Filter

Background

The Frost and Lee filters are based on models which do not use the statistical properties of the underlying scene.

In a joint study with CESR (Toulouse, France), CCRS participated in the development of the MAP Gamma Filter (Lopes, Touzi, Nezri and Low, IJRS, 1993).


Most well known Filters : The MAP Gamma Filter (cont’d)

Principle

The filter is based on the assumption that the (unspeckled) intensity of the underlying scene is gamma distributed.

The filter minimizes the loss of texture information better than the Frost and Lee filters within gamma distributed scenes.

It is suitable for a wide range of gamma distributed scenes, such as forested areas, agriculture areas, and oceans.

The filter preserves the observed pixel value for non-gamma distributed scenes.

See Figure 5.9 for the filter example.


Figure 5.9 - Map Gamma Filter


Original Image Map Gamma 11x11


Effects of Filtering

Whereas, adaptive filters (Lee, Frost and Gamma) preserve the mean value and are therefore preferable for SAR imagery (Figure 5.10).

Figure 5.11 shows that as the filter kernel size increases, so does the percent change in standard deviation.

A quantitative example of these effects on real data is shown in Figure 5.12.


Figure 5.10 - Effects of Filtering

Filter Size & Type vs % Change in Mean

Median 5x5

Filter Size & Type

Perc

enta

ge C

hang

e in

Mea

n

Median 7x7Median 3x3

Lee 5x5

Raw Lee 7x7Lee 3x3 Frost 3x3

Frost 7x7

Frost 5x5

Source: CCRS


Figure 5.11 - Effects of FilteringFilter Size & Type vs % Change in SD

Filter Size & Type

% C

hang

e in

Sta

ndar

d D

evia

tion

Raw Median 7x7 Lee 7x7 Frost 7x7

Frost 3x3

Lee 5x5

Lee 3x3Median 3x3

Median 5x5 Frost 5x5

Source: CCRS


Figure 5.12 - Effects of Filtering

Source: CCRS, Brown et al, 1993

Effects of Filtering on Sample Wheat Field Statistics, ERS-1 SAR

Mean Standard Deviation

% Change in Mean

% Change in SD Mean/SD

Raw

Median 3x3

Median 5.5

Median 7x7

Lee 3x3

Lee 5x5

Lee 7x7

Frost 3x3

Frost 5x5

Frost 7x7


Edge Detection in SAR ImagesApplication : Segmentation of the image into separate entities, classification Types of Edge Detection Filters:

Directional, Gradient, Laplacian, Sobel, Prewitt, Ratio Edge Detector

WarningsThe classical edge detectors (e.g. Gradient, Sobel) developed for imagery from optical sensors are not suitable for SAR images.Because of the multiplicative nature of speckle, they detect more false edges within brighter areas.Imagery must first be filtered (Gamma) prior to using the classical edge detectors.


Edge Detection in SAR Images (cont’d)

Potential alternatives

The ratio edge detector (R. Touzi et al., IEEE TGRS, 1988) is suitable for SAR images and does not require pre-filtering.

Performance of the ratio edge detector is better since information is lost during pre-filtering for the classical edge detectors.


Ratio Edge Detector Filter

(Touzi, et. al., 1998)

Original SAR image

Gradient image (5x5)

Ratio Edge Detector (5x5)

- For the gradient detector, the probability that a pixel of a homogeneousarea is assigned to edges (Pfa) is dependent on the mean power due to themultiplicative nature of the noise.

- The operator detects more false edges in brighter areas.

- The ratio edge detector is the ratio of the average of pixel values of twononoverlapping neighborhoods on opposite sides of the point.

- The Pfa does not depend on the mean power

- The performance of the ratio edge detector is a function of the size ofneighborhoods, the number of looks and the ratio of the mean powers.


The Touzi multi-resolution speckle FilterAll the most well known adaptive filters were developed under the assumption that the signal is stationary within the moving processing window of a fixed size (i.e. its mean and variance do not vary within the observation time).

The filters are not effective primarily when applied to fine structures such as roads and trails which are generally smoothed out by thefilters.

A new multi-resolution filter the Touzi Filter (Figures 5.13 and 5.14) was developed at CCRS (a part of PCI software 2002 version).

The size and the shape of the filter processing window are adapted to signal nonstationarity. The Touzi multi-resolution ratio edge detector is used for better filtering of contours and edges (Touzi et al., IEEE TGRS 1998)This permits more efficient speckle reduction and a better preservation of the scene spatial variations (texture, edges, point targets).

Source: R. Touzi, CEOS workshop 1999


Figure 5.13 - Touzi Filter


Original Image Touzi filter


Figure 5.14 - Touzi Filter

Original ImageTouzi filter

15X15

Lee filter7X7

RADARSAT-1 imageFine Mode


Introduction to Texture

Texture is the spatial variation of tones in an image.

Image texture may be qualitatively described as having properties like fineness, coarseness, smoothness, granulation, randomness, lineation, mottled, irregular, hummocky (Figure 5.15).

In a SAR image, texture has two components: (1) spatial variability in the scattering properties of the scene and (2) speckle.


Figure 5.15 - Image Texture

Corn Field Forest

300 m

Spatially Uniform TargetFine Texture

Spatially Non-Uniform TargetCoarse Texture

300 m

Source: Ulaby and Dobson, 1989


Texture Analysis

TextureTextural features contain information about the spatial distribution of tonal variations.Methods available:

Co-occurrence matrix (GLCM)Grey level difference vector (GLDV)Lacunarity (gap analysis)Neighbouring grey level dependence matrix (NGLDM)Spatial correlation functionModel-based approaches


Texture Analysis (cont’d)

Texture

Textural features statistics can be extracted using agrey level Co-Occurrence Matrix (GLCM).

User specific neighborhood parameters.

Examples of features from GLCM:

- Homogeneity - Mean - Contrast - Standard deviation - Dissimilarity - Entropy - Angular second moment - Correlation

Speckle suppression techniques may not preserve all scene texture details.


Contrast Stretch

A contrast stretch enhances visual interpretation (Figure 5.16).Matches data’s dynamic range to dynamic range of display.Involves the construction of a look-up table (LUT).LUT is a graphical model of the mathematical function selected.


Figure 5.16 - Contrast Stretch

Original image Linear Stretch

Rosario, Argentina


Linear Stretch

Effective upper and lower cutoff values are established.

Upper and lower histogram values are set to maximum & minimum limits respectively.

May use full or piecewise stretch.

Balance of the data are stretched linearly to fill the expanded display range.

See Figure 5.17.


Figure 5.17 - Linear Stretch


Nonlinear Enhancements

Distort the image radiometry.

Useful only for visual interpretation.

quantitative radiometric information can be lost.

spatial information is preserved.

results may not be replicable from scene to scene.


Histogram Stretch

Input display range may not be fully utilized.

Output display range makes full use of the dynamic range.

Enhances the contrast where frequency of occurrence is greatest.

Options include:- Inverse frequency- Frequency equalization- Gaussian normalization- Histogram matching


Inverse Frequency (or Infrequency)

Produce an image in which the bright pixels represent those grey levels in the original image which were infrequent.

LUT is derived from an inverted (upside down) histogram of the input image data values.

Useful for highlighting rare or small features in an image (lineaments or edges).

Figure 5.18 is an example of infrequency enhancement.


Figure 5.18 - Inverse Frequency Enhancement


Frequency Equalization

Redistribute pixel values so that there are approximately the same number of pixels for each data value available.

More for visual display than for image analysis.

Figure 5.19 is an example of Frequency Equalization.


Figure 5.19 - Frequency Equalization


Exponential Stretch

High-range brightness is enhanced and high histogram skew can be corrected.

Details in the higher part of the dynamic range are revealed.

An example of an algorithm for this stretch is ex.


Logarithmic Stretch

Low-range brightness is enhanced and histogram skew may be corrected.

Skewness is common and may invalidate some image analysis algorithms which assume a normal data distribution.

Also known as root Enhancement.

Root ( log N).

Tends to lend an overall brightening to the resultant image (see figure 5.20).


Figure 5.20 - Logarithmic Stretch


Power Law Stretch

Changes the image brightness, S, as a power law:

Snew = Sn

n > 1 enhances strong returns at the expense of weak returns.

n < 1 ( n ) enhances weak returns at the expense of strong returns.

The special case n = 2 converts a magnitude image to a power image.

Alters the probability distribution (histogram) of the data and may invalidate processes based on Gaussian assumptions.


CONVERSION FROM DN TO:

σ° or β°(dB)

σ° or β°(power)

INTERFEROMETRY- DEM generation- Coherence image- Surface change detection

FILTER(speckle reduction)- Adaptive filters

- Non adaptive filters

STEREOSCOPY- DEM generation- Planimetric feature

extraction

CHANGE DETECTION(e.g. ratio, difference)

CALCULATION OFTARGET SIGNATURES

CONVERT POWERVALUES TO dB

e.g. σ° (dB) = 10 log10 ( X )

MODELLING- Theoretical backscatter- Geophysical parameters

extraction

TEXTURE ANALYSIS(input for classification) FILTER

(speckle reduction)- Adaptive filters- Non adaptive filters

ENHANCEMENT(for visual interpretation)- High pass filters- Low pass filters- FFT filters- Contrast stretch

GEOMETRIC CORRECTION- Ortho-rectification using DEM- Slant / ground range conversion- Polynomial transformation

DATA FUSION- RGB-IHS Colour Space- Principal Component

Analysis- Vector Overlay

CLASSIFICATION- Supervised- Unsupervised

ACCURACY ASSESSMENT

AUTOMATED FEATUREEXTRACTION- image thresholding- edge detection, lineaments- directional filters (Sobel, etc.,)

OTHER DATA- multitemporal SAR- optical RS- geophysical- Thematic polygons

or vectors (GIS)- etc.

QUANTITATIVEQUALITATIVE

“TYPICAL” SAR IMAGE PROCESSING METHODOLOGY

INFORMATIONEXTRACTION

- Valued-addedinformation map

AMPLITUDEDigital Number

(DN)

AMPLITUDE + PHASESingle Look Complex

(DNI + DNQ)

STEREOSCOPY- terrain interpretation

sec5 form r.ppt [read-only] - unisi.itgarzelli/documenti/07-sar-filtering-enhancement.pdf · canada...

Documents