Geoinformation, monitoring and assessment Environment

Posted December 1999

Inventory and monitoring of shrimp farms in Sri Lanka by ERS SAR data, Part 3

by
Carlo Travaglia
FAO Environment and Natural Resources Service
James McDaid Kapetsky
FAO Inland Water Resources and Aquaculture Service
Giuliana Profeti
Consultant, Digital Image Processing
Environment and Natural Resources Working Paper No.1
FAO, Rome, 1999


For a printed copy of this publication, e-mail: Carlo.Travaglia@fao.org


< from Part 2

Speckle removal

SAR images are affected by a kind of noise called speckle. The speckle causes randomly scattered pixels to have particularly high or low values, thus increasing the error in classifying low-reflecting surfaces as water-covered surfaces.

The speckle is minimized by applying speckle suppression filters to the image. A filter is a matrix of values (also called template, box, window or kernel), whose dimensions are chosen by the operator. The filter matrix is moved over the image row by row and column by column. The pixel values covered by the matrix at a particular position are then used to define a new value for the pixel corresponding to the central element of the matrix (Richards, 1993). The noise reduction is generally accompanied by a loss of details; therefore the choice of the filter depends both on the characteristics of the image and on the kind of subsequent analysis.

The aim of the present study is to identify shrimp ponds, which may be small in extension, thus the loss of detail must be kept to a minimum. Also, the presence of noise may be partially tolerated as part of the analysis is performed visually by an operator. Finally, the entire image processing procedure must be as simple as possible, including the noise reduction technique, to allow for its use by people not highly trained in SAR image analysis and interpretation.

In order to choose the speckle removal procedure that satisfies those requirements, the following speckle suppression filters provided by ERDAS IMAGINE (version 8.3) have been tested:

These filters are generally applied iteratively until the desired effect is reached. Three combinations of filters have been applied to the three test sites and their results have been visually analyzed to choose the one which satisfies the balance between noise reduction and identification of the smallest shrimp ponds:

  1. Three applications of Lee-Sigma filter, increasing the window size (3x3, 5x5, 7x7) and increasing the coefficient of variation multiplier (0.5, 1, 2);
  2. Application of Lee and then Local Region filters at constant window size (5x5);
  3. Three applications of Frost filter, at increasing window sizes (3x3, 5x5, 7x7).

Analyzing each passage of the three filtering sequences on the test sites it has been noted that the smallest shrimp ponds are still visible only after a single application of the Frost filter, using a 3x3 moving matrix. Therefore the three images of the study area have been subjected to this filter to conclude the preprocessing procedure.

The images obtained from the sequential application of the Frost filter to the Seguwantiyu test site are shown as example in Fig. 17.

2.4 Classification

The purpose of this part of the analysis procedure is to identify water bodies, and possibly dykes, in the ERS SAR images. The identification of other types of land cover is not required. Thus, the images can be analyzed only by means of unsupervised classification procedures or by histogram thresholding.

As described in the preceding paragraph, water bodies are characterized by typically low values in SAR images. Conversely, the values of dykes surrounding shrimp ponds vary in SAR images according to their position relative to the satellite cross-track direction. Thus, dykes perpendicular to the cross-track direction have typically high values; dykes parallel to the cross-track direction are not visible in the image, and dykes at intermediate angles have values similar to those of other surfaces.


a) Original image

b) First application of Frost filter, 3x3

c) Second application of Frost filter, 5x5

d) Third application of Frost filter, 7x7

Fig. 17. Sequential application of Frost filter to the image of the Seguwantiyu test site

The analysis procedures are thus expected to identify water bodies and high-reflective surfaces, including some dykes. Other kinds of land cover are not of interest and the procedures assign all of them to another class.

The images acquired on 18 April 1996 on the three test sites and filtered with one passage of 3x3 Frost filter have been first classified using the ISODATA unsupervised classification procedure (Lillesand and Kiefer, 1987). The operator is required to specify the number of classes to identify (three in this case) and input the coastline that defines the area where the procedure is applied. After, the operator must visualize the results and identify the surface covers corresponding to the classes obtained from the procedure.

The procedure is thus almost completely automatic. The classifications obtained for the three test sites have been compared with the available ground truth. The unsupervised classification performs satisfactorily; for example, see the Uppadaluwa classified image in Fig. 18.


a) ERS SAR image 18/04/96, filtered with 3x3 Frost filter


b) Unsupervised classification of image a)

Fig. 18 Mapping of water bodies, Uppadaluwa test site

The second analysis procedure that may be applied to SAR images to identify water bodies and high-reflective surfaces is named histogram thresholding. It is a supervised procedure: the operator must analyse the image histogram to identify the peaks corresponding to water, high-reflective surfaces and other surfaces, and define the threshold digital values that separate them. All the image values are thus assigned to one of the three classes comparing their values with the threshold values. The operator is thus able to guide the assignment until a satisfactory result is obtained. The thresholding procedure has been applied to the same images of the three test sites. The results are almost identical to those obtained from the unsupervised classification. As thresholding requires a greater amount of operatorís time, it is suggested to use the unsupervised classification instead.

2.5 Boundary detection

The dykes surrounding shrimp farms may also be identified applying edge detection filters on ERS SAR images. These filters have the purpose of identifying the boundaries between homogeneous areas; the other information is lost in the output image. The Sobel filter (Richards, 1993) has been chosen for this study.

This filter is a non-directional operator that simultaneously calculates the horizontal and vertical gradient in the portion of the image covered by the filter kernel.

The result is equivalent to the simultaneous application of two directional kernels (Fig. 19 a and b).

These kernels may also be applied singularly to detect edges in horizontal and vertical directions. The other two kernels can be defined in analogy with these, to detect edges in diagonal directions (Fig. 19 c and d).


Fig. 19. Directional Sobel filters

Fig. 20. Uppadaluwa test site. Sobel filters applied to the 18/04/96 ERS SAR image. S-diagonal edges are displayed in red, vertical edges in green, and horizontal edges in blue

The application of each filter produces an output image that contains only the edges, defined by lines two pixels wide. A color combination of the filtered images allows an enhanced visualization of the boundaries in the study area (Fig. 20).

2.6 Proximity analysis

The occurrence of highly reflective surfaces around water surfaces is an indication of the presence of shrimp farms. The proximity analysis examines the boundaries of water bodies obtained from the unsupervised classification, up to a user-specified distance, to locate both highly reflective surfaces in the classified image and edges in the Sobel filtered images. The proximity analysis produces two "summary images" that synthesize the shrimp ponds-related information contained in an ERS SAR image.

The summary images allow the operator to locate the areas where there is a greater evidence of the occurrence of shrimp farms, and help in tracing the farmsí boundaries.

Fig. 21 shows the overlap of the two summary images for the Uppadaluwa test site.


Fig. 21. Uppadaluwa test site. Overlap of the two summary images and shrimp farms map obtained by visual interpretation (red line). The coastline is traced in brown

Highly reflective surfaces and sharp boundaries are displayed in black tones, water bodies are blue.

This figure shows also how the areas displayed in the summary images have been further analyzed using the criteria outlined in the first paragraphs of this chapter. The cluster of small water bodies located in the lower right corner of the image has been identified as rice paddies. The identification of the larger group of water bodies as shrimp farms has been confirmed and its contour traced on the SAR image using the summary image as reference, and completing it visually with the addition of a few areas of smaller ponds which had not been enhanced by the automatic procedure.

2.7 Field verification

Field verification was carried out by a four-person team, including the second author, in December 1998. The basic strategy was to verify the four classes produced by interpretation of ERS SAR data, reported on maps at a scale of 1:50 000: shrimp farms existing up to 18 April 1996, shrimp farms constructed between the former date and 16 October 1998, areas tentatively identified as shrimp farms and inland water bodies. In order to cover as much of the area of interest as possible, verification sites were selected that were adjacent to main roads. In order to discriminate between shrimp farms constructed before and after 18 April 1996, it was necessary to interview knowledgeable people.

At each verification site, a location (in both latitude/longitude and UTM coordinates) and estimated position error ("epe", in meters) were obtained using a GPS receiver. Verifications were carried out from the southern limit of the area of interest to the NE extreme and nearly to the NW extreme. In all, 32 waypoints were acquired. At some points where there was certainty of location, observations without GPS coordinates served as supplemental verification sites. The location of the sites field-checked, the estimated positioning error and the results of the verification are reported in Table 5.

The ground truthing indicated an 86 percent accuracy of the interpretation. The field observations permitted the interpretation keys to be refined and in this way some potential misinterpretation of the SAR data were eliminated. The accuracy of shrimp farms mapping, revised after the field verification, is, thus, estimated to be more than 90 percent.

Table 5. Location of waypoints and results of the verification
WP Latitude Longitude epe1 Image interpretation2 Ground truth

n.

deg

sec

d sec

deg

sec

d sec

m

18/04/96,03/07/98,16/10/98,05/03/99

december 1998

1

7

27

37

79

49

63

26

Reservoir.

Reservoir (Matha Weva).

2

7

28

39

79

49

71

52

No shrimp farms identified.

Rice paddies, forest.

3

7

28

61

79

49

81

32

Reservoir.

Reservoir (Tinabitiya Tank).

4

7

30

6

79

49

64

24

No shrimp farms identified.

Plantations.

5

7

29

80

79

49

45

44

Shrimp farms (1999).

Shrimp farms in construction.

6

7

29

89

79

49

0

27

East: Shrimp farms (1999).
West: Shrimp farms (1996).

East: shrimp farms in construction West: shrimp farms, built in 1995.

7

7

33

56

79

47

49

73

Shrimp farms (1996).

Shrimp farms, built before 1996.

8

7

32

67

79

47

71

40

Shrimp farms (1996).

Industrial shrimp farms.

9

7

30

92

79

47

98

33

Shrimp farms (1996).

Shrimp farms.

10

7

36

17

79

48

73

47

Shrimp farms (1998).

Trees and sandy terrain.

11

7

37

18

79

48

83

74

Shrimp farms (1996).

Shrimp farms, built before 1996.

12

7

38

31

79

48

58

41

No shrimp farms identified.

Rice paddies.

13

7

47

98

79

48

55

37

Shrimp farms (1996).

Shrimp farms, built before 1996.

14

7

39

58

79

48

11

38

Shrimp farms (1996 and 1998).

Shrimp farms.

15

8

13

52

79

45

21

23

Shrimp farms (1998).

Partially flooded vegetation, marshland, flooded fields.

16

8

11

11

79

44

49

26

Shrimp farms (1996).

Shrimp farms

17

8

10

56

79

44

52

23

Shrimp farms, uncertain assignment.

Abandoned shrimp farms, shrimp farms, bare and flooded areas

18

8

9

52

79

44

25

26

Shrimp farms (1996)

Shrimp farms, built in 1996

19

8

5

80

79

43

93

29

No shrimp farms identified.

Coconuts, mangroves, lagoon.

20

8

5

5

79

43

84

33

Shrimp farms (1996).

Shrimp farms

21

7

59

48

79

44

71

31

Shrimp farms (1996).

Shrimp farms.

22

7

59

19

79

44

94

53

Shrimp farms (1996).

Shrimp farms.

23

7

58

50

79

48

73

41

Shrimp farms (1996).

Shrimp farms.

24

7

58

20

79

48

70

30

Shrimp farms (1996 and 1998).

Shrimp farms.

25

7

58

79

79

49

35

28

Shrimp farms (1996), salt pans.

Shrimp farms, salt pans.

26

8

4

48

79

49

26

41

Shrimp farms (1996 and 1998).

Shrimp farms.

27

8

6

27

79

50

67

24

Shrimp farms (1998)

Semi-inundated area and farmland

28

8

6

88

79

50

26

37

Shrimp farms (1998).

Shrimp farms

29

8

6

53

79

50

15

51

Shrimp farms (1998).

Shrimp farms.

30

8

6

18

79

49

91

23

Shrimp farms (1998).

Shrimp farms.

31

7

52

82

79

48

93

25

Right side of the road: shrimp farms (1996).
Left side: shrimp farms (1996).

Right side of the road: vegetation.
Left side: shrimp farms.

32

7

47

93

79

49

24

41

Shrimp farms (1996).

Lagoon, coconut and rice paddies

1. epe = estimated positioning error; 2. Bold characters indicate interpretation errors

To: Inventory and monitoring of shrimp farms in Sri Lanka by ERS SAR data, Part 4



Back to Top FAO Homepage