PRACTICAL ON SAR IMAGE PROCESSING AND CLASSIFICATION

In this practical exercise you will learn the basic SAR image processing, such as speckle
suppression, edge detection, image enhancement, texture analysis and image
classification. You will be working on geocoded backscatter image of ENVISAT ASAR
IMS and APS (dated 16 Dec. 2007 and 21 Dec. 2007) of Manali sub-basin of Beas
river in Himachal Pradesh. All the analysis will be done in ERDAS Imagine software.

The fig. 1a and 1b shows the study area map in ETM+ and ASAR IMS. Use ERDAS
Viewer to see these images initially.

Fig. 1a: Landsat ETM+ FCC image of Manali sub-basin Fig. 1b: ASAR-IMS (C-VV) image of study area.

Part A: Speckle Suppression

All the SAR images needs to be corrected for speckle, before starting any feature
extraction or classification process.

Fig. 2a: Speckle suppression

window in RADAR module

You will execute speckle suppression filters available in ERDAS "RADAR->RADAR

INTERPRETOR->SPEKLE SUPPRESION" module fig.2 and select the filter which
gives best results for your application.
  Select the input file as "man16dec07vvdb.img" from your data directory. Before
starting any speckle suppression filters, check the option "Calculate Coefficient of
Variation" option and press OK (fig. 2b).

Fig. 2b: Calculation of the coefficient of variation in ERDAS and reading in sessions log.

 Go to the main ERDAS menu and select "session log" option and note down the
Coefficient of Variation value from session log file (fig. 2b). This initial estimate
will be used in the Lee, Lee sigma, frost and gamma map filters.
 Now select the filters: mean, median, lee sigma etc. and give unique output files
name for each filter output.

original ASAR Frost filtered

IMS IMAGE ASAR IMAGE
Fig. 2c: Original ASAR
image and speckle
suppressed images.

Lee sigma Local region

filtered IMAGE filtered IMAGE
Part B: Edge Detection/Enhancement
Edge enhancement is required in many applications such as automatic extraction of roads,
bridges, air fields, building polygons, forest and crop area boundaries, geological and
drainage features.
 select the edge enhancement from ERDAS Imagine options: "RADAR->RADAR
INTERPRETOR-> EDGE ENHANCEMENT" (fig. 3a).

Fig. 3a: Edge enhancement option in ERDAS and list of built filters.
 Select the input file initially as "man16dec07vvdb.img" (and later, select the
speckle suppressed images) for edge enhancement using the available filters such
as multilevel, Krisch etc. The filter's kernal details are also given (fig. 3a).Give
unique output files name for each edge enhanced filter output (fig.3b).

Original ASAR edges Lee Sigma filtered ASAR edges

Fig.3b: original
& edge enhanced
ASAR images

Multi level Weighted line

Part C: Image Fusion:

Image fusion for the current exercise is of 'data level fusion'. First you will do the image
fusion of ASAR HH, HV, VV by using the layer stack option. Secondly you will fuse
ASAR VV data with multi spectral ETM data using IHS transformation.

 Select the "ERDAS->Image Interpreter->Utilities-> Layer Staking" option

and add the multi polarization (HH, HV, VV) data of ASAR (fig.4a).

Fig.4a: Layer stack window Fig.4b: RGB color composite (HH:HV:VV) of ASAR.

 Give the unique output file name and select ignore zero in stats option. Run
the model by pressing OK and use viewer to see the final image in RGB
mode (fig. 4b).
 Now select "ERDAS->RADAR->Image Enhancement->Sensor merge"
option. Add ASAR VV image in grey scale image box and ETM+ image in
multi-spectral image box (fig.4c).
 Use IHS transformation with replacing saturation option for image fusion and
give unique output file name and observe the new fused image FCC in viewer
(fig.4d). You can also try fusion with PCA and IHS with Intensity.

Fig.4c: Sensor merge window Fig.4d: RGB color composite (4:3:2) of ASAR.
Part D: Image Classification:
The speckle suppressed and image enhanced product will now be used for unsupervised
and supervised classification in ERDAS (or SARSCAPE if available).

 First unsupervised classification will be done using "ERDAS IMAGINE

->CLASSIFICATION ->UNSUPERVISED CLASSIFICATION" option. It uses
the ISODATA statistical image classifier for the making the clusters and making
the classified map. Initially use the fused image of ASAR IMS (16dec07 and Nov
2000 ETM+ image) "man16dec07_etm_intihs.img" and later use Polarimetric
ASAR image (21 dec07 HH, HV, and 16 dec07 VV) "21dec07hh_hv_vv.img" for
unsupervised classification. (fig. 5a).
 Give the unique output file name, select no. of classes as 6 and uncheck output
signature option, and select initializing option as automatic. (fig. 5a).

(Fig. 5a): Unsupervised

classification process for
fused and polarimetric
SAR images

 Keep no. of iterations as 6 and threshold convergence as 0.950. Finally press OK

and open viewer to see results. Change the colors of classified classes (use ETM+
image as guide for assigning class names and color) (fig.5b,c).

Fig. 5b:SAR+ETM fused unsupervised classified image Fig. 5c: ASAR polarimetric unsup. cls. image
THEORY: UNSUPERVISED CLASSIFICATION

ERDAS IMAGINE uses the ISODATA algorithm to perform an unsupervised

classification. ISODATA stands for "Iterative Self-Organizing Data Analysis
Technique." It is iterative in that it repeatedly performs an entire classification (outputting
a thematic raster layer) and recalculates statistics. "Self-Organizing" refers to the way in
which it locates the clusters that are inherent in the data (ERDAS online help 2009).

The ISODATA clustering method uses the minimum spectral distance formula to form
clusters. It begins with either arbitrary cluster means or means of an existing signature
set, and each time the clustering repeats, the means of these clusters are shifted. The new
cluster means are used for the next iteration.

The ISODATA utility repeats the clustering of the image until either:

 a maximum number of iterations has been performed, or

 a maximum percentage of unchanged pixels has been reached between two
iterations.

THEORY: SUPERVISED CLASSIFICATION

The decision rules for the supervised classification process are multi-level:
o non-parametric
o parametric
If you have both non-parametric and parametric signatures, then you must decide which
decision rules will apply for both situations.

Non-parametric
If the signature is non-parametric (i.e., feature space AOI), then the following decision
rules are offered:
o feature space
o parallelepiped
With non-parametric signatures, you must also decide the overlap rule and the
unclassified rule.
NOTE: All signatures have a non-parametric definition due to their parallelepiped boundaries.

Parametric
For parametric signatures the following decision rules are provided:
• maximum likelihood
• Mahalanobis distance
• minimum distance

The Supervised Classification utility outputs a thematic raster layer (.img extension) and
optionally a distance file (.img extension). The distance file can be used for post-
classification thresholding. (ERDAS online help 2009). The details of each rule/method
of image enhancement, fusion, supervised and supervised classification can be found in
Digital Image Processing book by Jensen.