2. Methodology

2.1 Data and software used

The study area is completely covered by two ERS-2 SAR images acquired on 2 December and 23 December 2002 respectively (Figure 2 and Table 3). The spatial resolution of ERS-2 SAR images is of 12.5 x 12.5 m. The two images were provided by the European Space Agency (ESA), in the context of their scientific research programme, in the Ellipsoid Geocoded Format (GEC). These images are system and ground range corrected, and were georeferenced and rectified into the Universal Transverse Mercator Projection (ellipsoid WGS84, zone 51 N). They have not been corrected for terrain distortion, as this was not necessary, the aquaculture and fisheries structures occuring in flat areas.

The two images were specifically acquired by ESA for this study, by selecting two acquisitions made during descending and ascending orbits with the least possible time interval in-between.

Orbit direction during the acquisition is extremely relevant because in descending orbits the scanning direction of the sensor is approximately opposite to that in ascending orbits. This in turn influences the characteristics of the SAR images, in which features are enhanced in a complementary way, as described in section 2.3.

TABLE 3
Characteristics of the satellite data used

¹ Ellipsoid Geocoded Image. ² SAR Georeferenced Fine resolution. ³ Floating along-track between two frames. UL = Upper Left, UR = Upper Right, LR = Lower Right, LL = Lower Left.

A RADARSAT-1 SAR image acquired on 4 February 2001 in the Georeferenced Fine Resolution (SGF) format was purchased from RADARSAT International. Its ground resolution is of 6.25 x 6.25 m. This image is also ground range and system corrected, and has been georeferenced and rectified into the geographic system, ellipsoid WGS72. It does not cover the entire study area, but includes the zones where the majority of the aquaculture and fisheries structures are located (Figure 2 and Table 3).

In the study of coastal features and of some aquaculture and fisheries structures, the tide stage at the time of acquisition of the radar data may be of interest. For some considerations, images should be acquired at high tide, in order to delineate the land that is submerged only in exceptionally high tides. This would allow to reduce uncertainties in the visual interpretation stage, as coral reefs, sand bars and other coastal features are submerged. The coastline charted in the topographic maps is usually derived at the average high tide level and, therefore, could be directly compared with the one obtained from the images. On the other hand, radar data acquired at low tide would definitely enhance the possibility of mapping fish traps, the surface reflecting the radar beams being greater.

Unfortunately, the only tide measuring station in the study area is in San Fernando, which is located almost outside Lingayen Gulf. Due to the conformation of the gulf, the tidal range is greatest along the coast inside it, and particularly in its southernmost part. Thus, the tide stage cannot be derived from the records obtained in San Fernando, and consequently it has not been taken into account in the selection of the images.

Hence, the choice of the image acquisition period was based only on the season. All images were acquired in winter, during the dry season, when rice fields are not flooded. This allows to minimize the errors in the visual interpretation of the images.

The most detailed reference cartography available for the study area is a series of topographic maps at 1:50 000 scale, published in 1977. Data are mapped into the geographic system, ellipsoid Clarke 1866, datum Luzon. The study area is covered by ten topographic maps (Figure 6). The maps were scanned and then the features of interest were digitized using the GIS software Arc View 3.2, as described in section 2.2.

FIGURE 6
Topographic maps, 1:50 000 scale, covering the study area

All the satellite images acquired for the present study were pre-processed and reprojected into the cartographic reference projection using the software Erdas IMAGINE 8.5. Satellite imagery and vector data were analyzed jointly in the visual interpretation stage using the GIS software Arc View 3.2.

2.2 Preparation of the vector database

The objective of the visual interpretation is to identify and map aquaculture and fisheries structures in the SAR images of the study area. These must be then compared with the cartographic data. The best strategy to perform this task is to create or transform all data in vector format, and analyse them using a GIS software.

For this reason, the first part of the preparatory work consisted in obtaining a vector database from the topographic maps of 1977, containing all the data relevant for this study. The maps were scanned at 300 dpi and the raster data thus obtained were geocoded by applying the rubber sheeting resampling procedure. This procedure puts each reference point at the given coordinates, and uses a linear interpolation to reproject the other elements of the grid. In this case, the reference points were the corners and all the grid intersections of each map. The vector information was then created by on-screen digitizing of the displayed raster maps. The scale of the topographic maps is 1:50 000, so to minimize errors the display scale for digitalization was 1:25 000 or larger.

Among the data reported on the topographic maps, only those relevant in the analysis of aquaculture and fisheries structures were digitized. Two layers were created in ArcView shapefile format, containing respectively polygons and polylines; in each case, the corresponding land cover and land use is specified by an attribute code.

The polygon layer contains the following classes:

Fishponds, empty fishponds, fishponds with nipa, and nipa

All these land use classes are directly relevant for the present study; the corresponding areas were digitized. In particular, adjacent groups of fishponds of the same category were digitized into the same polygon if they appear to be separated only by dykes, and as separate polygons otherwise.

A different code was assigned to the polygons of the four categories, as the related information might be useful in the subsequent analysis. As fishponds are periodically emptied for maintenance purposes, their being empty at the moment of the aerial survey upon which the maps are based does not necessarily imply that they were abandoned.

Regarding fishponds containing nipa palms (Nypa fruticans, Wurmb), this plant is extensively used as roofing material. Thus, it is sometimes grown inside fish ponds to increase their total revenue. Nipa palms emerging from water interact with the radar signal increasing the backscatter intensity over the pond surface; thus, data on their presence may be useful in the visual stage. For the same reason, nipa palms growing outside fishponds were mapped as well.

Mangroves

This land cover class is scarcely present in the study area.

Large rivers and lakes

Wide rivers can be easily identified in the SAR images. The knowledge of their boundary assists the interpreter in analysing the presence of fisheries structures inside the river itself, if its position and shape has not changed in time. In the study area, three of the largest rivers (Agno, Panto and Cayanga) actually show major modifications in the SAR data with respect to their position and extension reported in the topographic maps of 1977. The knowledge of the location and extension of natural lakes is also useful to avoid misinterpretation errors.

Actually, lakes are easily separated from fishponds as the latter are much more regular in shape and bounded by easily discerned dykes; however when the water-covered surface has a small extension, the dykes and the shape are less evident and errors may arise.

Salt pans

They cannot be separated from fishponds when they are flooded, thus knowing their location is useful to avoid interpretation errors.

Coral reefs, sand banks

Fishponds and other aquaculture structures may be built on open, shallow waters. Coral reefs have sometimes been destroyed to accommodate them, hence these areas were digitized to analyse whether this has occurred in the study area. On the contrary, sand banks are generally avoided as fishponds location because they correspond to areas where highly dynamic processes of sedimentation/erosion occur.

Mainland, islands

The coastline has been digitized as well; islands are included only if located in the open sea. The coastline on the maps corresponds to the mean lower low water. This layer has been compared with the coastline obtained from the 2001-2002 images, to enhance modifications that impacted on the location of aquaculture and fisheries structures. Digitizing of the coastline using the SAR images is described in section 2.5.

The polyline layer obtained from the topographic maps contains the following classes:

Roads and railroads

Includes the railroad and all types of roads; does not include tracks and trails. All roads have the same code. Roads and railways are relevant in the study of fishponds as they act as constraints for their expansion, facilitating at the same time fish transport to markets.

Rivers

Includes rivers whose width is too narrow to be represented in scale. The river network is an important element in the visual interpretation stage: fishponds must be connected to flowing water, and thus the ponds are always located in proximity of the main river network.

Figure 7 shows an example of the vector data obtained from the topographic maps.

FIGURE 7
A portion of Dagupan City topographic map (sheet 7074 II) and the derived vector data

2.3 Mapping aquaculture and fisheries structures by satellite imaging radar

Radar is the acronym of Radio Detection and Ranging. An imaging radar is an active device that transmits microwave pulses toward the Earth surface and measures the magnitude of the signal scattered back towards it. The return signals from different portions of the ground surface are combined to form an image. A Synthetic Aperture Radar (SAR) is a special type of imaging radar. It is a complex system that measures both the amplitude and phase of the return signals. Their analysis exploits the Doppler effect created by the motion of the spacecraft with respect to the imaged surface to achieve high ground resolution.

As the source of the electromagnetic radiation used to sense the Earth surface is the system itself, it can be operated during day and night. The atmospheric transmittance in the microwave interval used by remote sensing SAR systems is higher than 90 percent, even in the presence of ice and rain droplets (except under heavy tropical thunderstorms); thus, SAR can acquire data in all weather conditions.

A drawback of SAR imaging is the presence of noise (speckle) in the images. The noise is created by constructive and destructive interference between the backscattered energy from different portions of the ground surface included in the same cell (pixel) of the SAR image. The value of the pixel is thus increased or decreased; the SAR image appears to be covered by randomly scattered bright and dark spots.

In all satellite imaging SAR systems, the pulses are emitted sideways, downwards to the Earth’s surface and perpendicularly to the flight direction. The ERS SAR acquires strips of imagery approximately 100 km wide, 250 km to the right of the sub-satellite track; the incidence angle of the emitted pulses in the middle of the imaged stripe is 23°, and the ground resolution is of 12.5 x 12.5 m.

The SAR on board RADARSAT operates in several acquisition modes, obtaining images with varying resolution and size. To detect aquaculture and fisheries structures the highest possible resolution is necessary; this corresponds to the Fine Resolution acquisition mode. The ground resolution is 6.25 m; the acquired image is 50 km wide, it is located 385 km to the right of the sub-satellite track, and its mid-image incidence angle is of 44.259°.

Aquaculture structures are evident in SAR data because their components influence in a peculiar way the radar backscatter.

An analysis of fishpond appearance on SAR data has already been conducted by Travaglia, Kapetsky and Profeti (1999). Fishponds are small enclaves of calm water surrounded by dykes on all sides. A dyke is an earthen wall whose thickness ranges approximately from half a metre to several metres, and whose elevation from the water surface is at the most a metre. While a calm water surface behaves like a specular reflector, sending only a small part of the signal back to the sensor, a dyke reflects back a large amount of the incoming energy, because its sides intersect the surrounding water at approximately a right angle, creating a “corner reflector” (Figure 8).

FIGURE 8
Interaction of radar beams with dykes and water surfaces on a group of fishponds

As shown in the figure, the radar signal bounces off of both planar surfaces and is reflected directly back toward the antenna. The pixel corresponding to the corner reflector has a high value, and thus fishponds appear as dark areas surrounded by bright, elongated structures.

Due to the peculiar acquisition geometry, the position of a corner reflector in a SAR image does not correspond exactly to the orthogonal projection of the dykes on a map. The position of the dykes in the image is shifted along the scanning (cross-track) direction and toward the sensor, and their apparent extension does not correspond to the actual one. Actually, the SAR signal of each pixel is obtained by averaging the radiation reflected back by the various surfaces inside the area corresponding to the pixel. Thus, if the imaged area contains a small but highly reflective object, the average backscattered radiation is almost equal to the single contribution of the object. As a result, the pixel assumes a high value and the dyke appears to be as big as the entire pixel. The multiple reflection on the dyke may also spread the high return signal to the surrounding pixels, increasing their values as well. Therefore, the extension of a dyke may appear larger in a SAR image than in reality.

The return signal of elongated objects varies also as a function of the angle between the object and the cross-track direction (Figure 9). Surface features oriented in a parallel way with respect to the scanning direction are less evident than those oriented perpendicularly to the scanning direction. Hence, if a dyke is parallel to the cross-track direction, it may escape detection.

FIGURE 9
Return signal as a function of the angle between a dyke and the cross-track direction

The ERS-2 satellite follows a quasi-polar orbit, and as described previously its scanning direction (or cross-track direction) is right of the sub-satellite track. Thus, in descending orbits (from the North Pole downwards) the scanning direction is approximately opposite to that in ascending orbits (Figure 10). Consequently, surface features are highlighted in a different, complementary way on a pair of images acquired respectively in ascending and descending orbits.

The angle between the scanning direction of the two ERS SAR images used in this study is of 152.708 degrees. A comparative analysis of both images allows to identify properly all features, if they are acquired at a short time interval in order to minimize changes over the imaged surfaces between the two acquisitions.

FIGURE 10
The angle between the scanning directions in the SAR data used in the study

The other aquaculture and fisheries structures influence the radar signal in a similar way. The vertical sides of fish cages, pens and traps, emerging from the water surface, create the corner reflector effect that allows to identify them. For example, Figure 11 shows the interaction of SAR pulses with a fish cage. The sides of the cage oriented perpendicularly to the scanning direction are brighter in the SAR image.

In the smaller cages, the extension of the water surface inside is very small with respect to the sensor resolution and may not be represented in the image. As a result, the cage will appear as a bright group of pixels on the dark sea surface (Figure 12). The same happens to the smaller fish pens.

Both ERS and RADARSAT SAR sensors operate in the C-band (frequency 5.3 GHz, wavelength 5.6 cm). ASAR system generally sends out either horizontally (H) or vertically (V) polarized pulses, and collects either horizontally or vertically polarized return signals. The ERS SAR sends and receives vertically polarized signals (VV), while RADARSAT SAR sends and receives horizontally polarized signals (HH). Thus, both these sensors measure the portion of the backscattered signal which has maintained the original polarization.

FIGURE 11
Interactions of radar beams with a fish cage

The differences in the appearance of various coastal land features, including fishponds, on ERS and RADARSAT imagery were studied by Paringit et al. (1998) on the area of the Panay-Guimaras Strait (the Philippines) using the airborne NASA/JPL AirSAR polarimetric system. Their results show that the mean backscattering coefficient of fishponds is slightly higher on C-band images acquired in VV polarization than in HH. VV data are, however, more sensitive to sea surface roughness (Touzi, 1999) which in turn depends on wind speed. Wind speeds greater than approximately 1.5 m/s (Fingas and Brown, 2000) create waves that increase the return signal intensity in C-band, diminishing the contrast among sea surface and the structures located offshore. The contrast keeps diminishing as the wind speed becomes more intense. This effect is clearly shown in Figure 12.

Fish cages are evident in the RADARSAT-1 image and in the first ERS image (of 2 December 2002); they are barely visible in the second ERS image (of 23 December 2002) due to the increased sea surface roughness. It should be noted that in the second ERS image the coastline appears different, as the image was acquired during the receding tide; the emerging coral reefs contribute to increase sea surface roughness, thus reducing the possibilities of detecting fisheries structures.

Finally, the appearance of the structures in the SAR imagery is also greatly influenced by the spatial resolution of the sensor. Fisheries structures are generally made out of thin components and cover limited extensions; thus, the highest the spatial resolution the higher the possibility of detecting them. In particular, the smaller fish pens may not be evident in ERS SAR images, and fish traps are generally too thin to be detected; they appear only in the higher-resolution RADARSAT image (Figure 12).

FIGURE 12
Sea state and coastal aquaculture and fisheries structures mapping

The visual interpretation procedures used to map coastal aquaculture and fisheries structures are described in section 2.6.

2.4 Image pre-processing procedure

In order to perform the visual interpretation of the SAR images, they must all be geocoded in the same projection of the reference cartography. Speckle-reducing filters also were applied to the images to verify whether it was possible to enhance their interpretability.

The SAR images were provided already geocoded: the ERS-2 projected into UTM/WGS 84, and the RADARSAT-1 into geographic/WGS72, as described in section 2.1. The reference cartography was represented into geographic/Clarke 1866, thus all the images were reprojected into geographic/Clarke 1866 so that they could be overlaid with one another and with the cartography.

The automatic reprojection procedures provided by the Erdas IMAGINE software were applied at first, but when the images generated by these procedures were overlaid on the cartography they showed consistent misplacements. It has thus been necessary to manually geocode each image, using the “non-linear rubber sheeting” procedure. This method is based on the identification, over the image and over the cartography, of a large number of ground control points (GCPs). The coordinates of each pixel on the new geocoded image are then obtained interpolating nonlinearly the coordinates of the surrounding reference points. All the three SAR images were geocoded using more than four hundred GCPs, reaching a RMS error lower than 1.5 pixels; the output pixel size is 0.00005825 decimal degrees, equivalent to 6.25 m.

SAR images are affected by the presence of noise (speckle), created by constructive and destructive interference between the backscattered energy from different portions of the ground surface included in the same pixel of the SAR image. The value of the affected pixels is thus increased or decreased; the SAR image appears to be covered by randomly scattered bright and dark spots.

Thus, to complete the image preparation, it may be useful to apply speckle reducing procedures to the SAR images in order to increase their interpretability.

A simple, yet useful technique has been tested by Profeti, Travaglia and Carlà (2003) on multi-temporal SAR data to improve the visual interpretation of fish ponds. This technique enhances time-invariant spatial features and reduces speckle, without compromising the geometrical resolution of the images. This method allowed to obtain good results; however, it can be applied only on multiple images of the same area acquired by the same sensor in the same acquisition geometry, while in this study the two ERS-2 images were acquired in ascending and descending orbits; therefore, it is not applicable in this case.

The most common speckle removal procedures are based on adaptive spatial filtering based on local statistics. The filters analyse each pixel’s contextual information and produce a new image in which the value of each pixel is obtained from the values of its neighbouring pixels in the original image. Regardless of the specific filtering technique, noise reduction is achieved at the expense of the geometric detail of the image. Several filters proposed in literature (Lee, 1980 and 1981; Frost et al.,1982; Li, 1988) were tested upon each type of fisheries structure to be identified in the image, to evaluate their effectiveness in improving the structures’ visual appearance.

The analysis of the results shows that the original images are sharper and richer in details, very useful for visual interpretation purposes. Consequently, no speckle removal filters were applied.

2.5 Digitalization of the shoreline

Differences among the shoreline profile can be observed between the two ERS-2 images. On the first (acquired on 2 December 2002), the emerged land is wider than on the second (acquired on 23 December 2002) and part of the coral reef is also visible. The difference between land and water among the RADARSAT and the second ERS-2 image is small, and is probably more related to scale difference and geocoding than to tide stage.

At present, several different shoreline definitions are in use by various state and local authorities. The U.S. National Oceanic and Atmospheric Administration (NOAA) has adopted as standard shoreline the approximate line where the average high tide, known as Mean High Water (MHW), intersects the coast. In our case, the Philippines topographic maps use the Mean Lower Low Water (MLLW).

As no information on tides was available for the study area, it has not been possible to acquire images in a determined tide stage. Thus, it was decided to delineate the coastline by visual interpretation, using the image in which the coastline was more evident. To decide which image was best suited to be the reference for shoreline mapping, the scientific literature on this subject was reviewed. The use of airborne and spaceborne SAR imagery to delineate land boundaries has been tested widely in the last years; for example, RADARSAT imagery has been used to map the coastline on behalf the Digital Marine Resource Mapping (DMRM) program, initiated by the Government of Indonesia in 1996 (Hesselmans et al., 2000). A wider scientific research on the use of new technologies for shoreline mapping is being conducted by NOAA and the U.S. National Geodetic Survey on behalf the Coastal Mapping Program. It includes experiments on the use of satellite SAR imagery, whose results show that RADARSAT fine mode (HH) enables to map the coastline within 28 m and at 98 percent confidence level with respect to shoreline data produced using conventional photogrammetric processes (Tuell, Lucas and Graham, 1999). Other sources confirmed that HH imagery is better suited for shoreline mapping than VV imagery (although quadpol image data are considered to be the most suitable at all).

Therefore, the RADARSAT fine mode image has been used to map the coastline in the small portion of the study area it covers (Figure 2). The ERS-2 image acquired on 23 December 2002 has been used to complete the coastline at high tide, while the other ERS-2 SAR image (of 2 December 2002) has been used to map the low-tide boundary.

To map the low-tide coastline, the second ERS-2 image was overlaid with the land/sea boundary at high tide. Whenever the differences were wider than two pixels, they were mapped; this limit distance has been assumed sufficient for compensating geocoding errors and positioning errors related to the different scanning direction.

2.6 Mapping procedures

The description of the appearance of aquaculture and fisheries structures in SAR images, outlined in the previous sections, was used in the visual interpretation of the images.

The visual interpretation was performed using the Arc View software, as it is more suited for on-screen digitizing of the boundaries of the features. Two vector layers were created in order to collect the polygons and polylines of the classes of interest. Their content is described respectively in Tables 4a and 4b.

Polygons and polylines were digitized in the cartographic reference projection of the Philippines (section 2.1).

TABLE 4a
Classes identified in the SAR images: Polygon layer

For each element located in the images, the following parameters were calculated:

- Polygons: area (km²) and perimeter (km);
- Polylines: length (km).

The global area or length of the elements in each class have then been calculated and compared with the available ground truth from the topographic maps. The results are described in Chapter 3.

TABLE 4b
Classes identified in the SAR images: Polyline layer