4.4 Suitability of data to objectives

The data being used in the study are derived from satellites which were not designed specifically to monitor marine pollution. It is therefore important that the constraints and limitations of the instuments are understood in order that the information that is derived can be suitably qualified. The physics of the interaction of the electromagnetic radiation used by Clean Seas to monitor the sea surface is well understood but the actual performance of the real instruments operating in a real-world environment is not. The following sections detail a few of the limitations that must be taken into account when planning or operating a marine pollution monitoring system based on satellite data.

4.4.1 Wind speed dependency of SAR

It is a well-known phenomenon that the detectability of oceanic surface films by SAR sensors strongly depends on wind speed: at either very low wind speeds (below about 2 ms^-1) or very high wind speeds (above about 10 ms^-1) oceanic surface films cannot (or only hardly) be identified. Gade et al. (1998a,b) have recently shown that the reduction of the NRCS by marine surface films decreases with increasing wind speed, because at high wind speeds, wave breaking and the permanent action by the wind cause a reduced overall damping capability of the surface film.

In order to demonstrate this strong wind speed dependence, we have analysed a SAR image acquired over the North Sea test area on March 14, 1997, at 10:40 UTC (Figure 4-17). At the time of the image acquisition, an atmospheric front passed the area, so that the wind speed strongly increased (which in turn results in different intensities of the radar backscatter on both sides of the front, see Figure 4-17a).

Figure 4-17. a) ERS-2 SAR image of the North Sea west of the West Friesian Islands (orbit 10068, frame 2529, 24-Mar-1997, 10.40 UTC, image dimensions 100kmx100, image centre at 53¼21'N, 4¼26'E) showing radar signatures of oil pollution and of an atmospheric front; b) section (30kmx30km) of the SAR image marked by the white square in a); c) image intentsity scan along the line shown in b). The right scale denotes the wind speed derived from the NRCS.

South-west of the front, surface films are visible as dark patches and lines, whereas north-east of the front no such signatures can be observed. A dark patch corresponding to a surface film is placed exactly on the front line (see the enlarged part of the SAR image shown in Figure 4-17b). In order to study the dependence of the radar signature of a surface film on wind speed we have calculated the reduction of the normalised radar backscattering cross section (NRCS) along the scan line inserted into Figure 4-17b. This scan (see Figure 4-17c) shows that on the low-wind side (towards point A), the NRCS reduction is about 10 dB (i.e., the NRCS drops down to the noise floor), whereas it is only about 6 dB on the high-wind side (towards point B, see the horizontal lines added into the scan). This finding is in accordance with recently presented results of field experiments in the North Sea (Gade et al., 1998b).

On the right-hand side of Figure 4-17c the wind speed is added which has been calculated from the NRCS values by using a wind scatterometer model for retrieving sea surface winds from ERS scatterometer data (CMOD-4 model (Stoffelen and Anderson, 1997)). Mean wind speeds derived from a numerical model driven by the Deutscher Wetterdienst (DWD; German weather service) are in good agreement with those retrieved from the SAR image (i.e., they lie between 1 and 2 ms^-1 south of the front and between 4 and 5 ms^-1 north of the front). Especially the wind scatterometer model seems to underestimate the wind speed in the slick-free areas of the water surface, since a minimum wind speed of about 2 ms^-1 is needed for the generation of the small-scale surface roughness which is responsible for the radar backscattering. However, from Figure 4-17c it is obvious that an additional underestimation of the wind speed is caused when the water surface is slick-covered, because of the additional reduction of the NRCS by the slick.

The same numerical model from the DWD was used to get an estimate of the influence of the mean local meteorological conditions (namely the wind speed) on the overall detectability of marine oil pollution. The limited area (or nested) model (Europa-Model) for the North Atlantic and Europe area with a mesh size of approx. 55 km is embedded in a global model with a mesh size of approx. 200 km. The model results for the three test areas, each for the entire period of SAR image acquisition within Clean Seas (December 1996 until November 1998), are shown in Figure 4-18. The upper row contains the values calculated for the summer periods (April — September) and the lower row contains the values calculated for the winter periods (October — March).

It can be seen that the maximum mean wind speed in the Baltic Sea test area (left column of Figure 4-18) lies between 8 and 9 ms^-1 during summer and between 9 and 10 ms^-1 during winter. The corresponding values for the North Sea test area (middle column) are 7 and 8 ms^-1 (summer) and 10 and 11 ms^-1 (winter), and for the Gulf of Lion test area they are 6 and 7 ms^-1 (summer) and 8 and 9 ms^-1 (winter), respectively. Thus, on average, oil spill detection using SAR techniques should be most successful in the Gulf of Lion during the summer period, and it should be least successful in the North Sea test area during winter period.

In Figure 4-20 the mean spill-covered water surface in the three test areas is shown. For the generation of these contour plots the areas of the detected oil pollution have been derived by taking into account those dark patches in the SAR images, where the radar backscattering is significantly reduced. The obtained values have then been weighted by the number of acquired SAR images of that particular area.

In all three regions we found higher oil pollution during summer months (upper row; April — September) than during winter months (lower row; October — March), which can be explained by the overall higher wind speeds in all test areas during winter time (see Figure 4-18). Moreover, the Gulf of Lion seems to be the test area with highest detected oil pollution, whereas the pollution is least in the Baltic Sea. Taking into account the results from the DWD model, this might be caused by the difference in mean wind speed, which in turn causes a different visibility of oil pollution in SAR images. However, it is of course also possible that the Gulf of Lion is simply the most polluted test area. Especially the sea area off the mouth of the river Llobregat, close to Barcelona, seems to be permanently polluted because of the river runoff (which in turn seems to be polluted by local industry). In this particular area a river plume was detected in all analysed SAR images.

In order to improve our statistical analysis we included the mean local wind-speed, as derived from interpolated values of the DWD model. As a first step we calculated a histogram showing the distribution of the detected oil spills as function of wind speed. As shown in the left panel of Figure 4-19 (blue histogram) we detected most oil spills at mean (modelled) wind speeds between 3 ms^-1 and 4 ms^-1. The middle panel of Figure 4-19 shows the wind speed distribution of the DWD model with a maximum between 5 ms^-1 and 6 ms^-1 (purple histogram).

Figure 4-19. Left: histograms of the distribution of detected oil pollution as function of wind speed, middle: distribution of the DWD model winds, right: 'normalised oil spill visibility' calculated as the ratio of the histograms for oil spills and model winds.

Figure 4-20. Mean slick-covered water surface for the three Clean Seas test areas, as obtained from the analyses of ERS-2 SAR images. The upper row contains data for the summer periods (April-September) and the lower row those for the winter periods (October-March).

The right panel of Figure 4-19 shows the "normalised oil spill visibility" (NOSV) calculated as the (normalised) ratio of the two above (red histogram). That is, the NOSV gives a better estimate of the detectability of marine oil pollution, independently of the local wind conditions of this particular study. From the right panel of Figure 4-19 we see that in general higher wind speeds cause lower detectability, which has already been known. However, for the first time, our analysis gives a quantitative result of the maximum (model) wind speed where oil spill detection in European coastal waters is possible. In particular, at wind speeds well below 10 ms^-1 oil spills are detectable, whereas at 10 ms^-1 wind speed, and even more, the definite detection of marine oil pollution seems to be almost impossible. This results is even more important, since the mean wind speed in the central North Sea during winter time is above 10 ms^-1, see Figure 4-18.

Finally, we have generated plots showing the mean oil-spill coverage in the Clean Seas test areas, but taking into account only those SAR images, which have been acquired when the local wind speed was 9 ms^-1 or below. By doing so we are neglecting SAR images acquired at high wind speed conditions (above 9 ms^-1), thus yielding a bias of the obtained results. In Figure 4-21 the respective contour plots are shown.

By comparing the respective panels in Figure 4-20 and Figure 4-21, one sees that neglecting SAR images acquired at high wind speeds does not yield significant changes for the summer periods (upper rows) in all test areas. However, especially in the North Sea and Baltic Sea (left and middle columns), the neglect of the high-wind SAR images yields significantly higher values of the mean oil-covered water area. That is, our improved statistics show that the north-western Mediterranean Sea seems not to be the test area which is by far most polluted, but that similar quantities have been detected in all three Clean Seas test areas. However, in order to obtain more reliable statistics, a greater number of SAR images from the same test areas must be analysed.

Parts of the above results have been presented at the EARSeL 1998 Symposium (Enschede, The Netherlands (Gade et al., 1999)), the International Symposium on Marine Pollution 1998 (Monaco), the IGARSS conferences 1998 (Seattle, USA (Gade and Ufermann, 1998)) and 1999 (Hamburg, Germany (Gade and Redondo, 1999a)), and at the Oceans 1999 (Seattle, USA (Gade and Redondo, 1999b)). Furthermore, they are included in a peer-reviewed paper to be published in Science of the Total Environment (Gade and Alpers, 1999).

One of the major limitations for using thermal or optical sensors over European waters is the high percentage of cloud cover. The cloud cover was found to be a restriction in all three test areas, but in particular the Baltic and North Sea regions. The statistics of cloud cover obtained from the ATSR data processing scheme shows the overall extent of total cloud cover during the Clean Seas project.

Table 4-2. Number of ATSR scenes not totally cloud covered.

For the Gulf of Lion, almost 80% of the GSST scenes received had some cloud free information (Table 4-2). In general, the GBT scenes were more often cloudy, this is due to the GBT scenes principally being from the November 1996 - April 1997 period, when all the regions were more cloudy, whereas the GSST scenes cover summer periods as well.

However, for tracking features, more significant cloud free areas are required, and Table 4-3 shows the number of scenes during 1997 that were more than 50% cloud free for each region (a complete year has been shown to reduce seasonal bias in the statistics). Also shown is the average percentage of good data in an image for each region during 1997. The Gulf of Lion has almost 25% good data on average, making it a far simpler area to work with infrared and visible data than either the Baltic or the North Sea regions.

Table 4-3. Number of ATSR GSST scenes more than 50%cloud free 1997, and average percentage good data during 1997.

More important than the absolute amount of cloud in infrared images is the consistency of cloud coverage from one image to the next. The time series of SeaWiFS images from the May 1998 test period (Figure 4-22) for the North Sea shows this problem well. Although the time period is generally fairly cloud free, there is still a significant amount of cloud in each image, and tracking features from one image to the next is complicated by the moving cloud. The method of front tracing adopted by Ecole des Mines goes some way towards addressing the problems of partial cloud cover in AVHRR images when studying frontal regions, but does not help when attempting to obtain information such as sediment load from partially cloudy areas.

An additional problem was encountered in the North Sea region, where the area of interest included very near-shore waters. Here, coastal fog could often occur, obscuring the infrared signal within 5-10 km or the shore. Figure 4-1 shows one example of this, where the rest of the image is almost cloud free, but very close to the Dutch coast is totally obscured by fog.

In studies such as the Clean Seas test cases, use of remote sensed data is attempting to gain insight into the imaging mechanisms of the array of different sensors available on currently operational satellites. One important method of approaching this is to use synchronous, or near synchronous, images from different sensors to "add information" from one sensor to another. As there are only a limited number of cases where (e.g.) ATSR and SeaWiFS image the same area of sea within a few hours of each other, cloud cover can add another, yet more restrictive limitation. As atmospheric processes can occur at time scales as short as an hour, or even less, changes in cloud cover between images taken on the same day may remove any possible data pair matches. The Case study for 15 July 1997 in the Baltic [section 5.1] illustrates this problem where some of the features can only be seen on one or two of the infrared and visible images due to changes in the cloud cover over a time span of only a few hours.

It was unfortunate that there were not more data available from the OCTS sensor, flown on board the Japanese ADEOS satellite, as this sensor measures both infrared and visible at the same time. This alleviates the problem of attempting to get both infrared and colour from non-coincident sensors. The launch of ADEOS-2 (expected in 2002) should give a second opportunity to test out this sensor.

The limitations imposed by using infrared and visible data in such cloudy regions of the world emphasises the importance of using modelling to integrate data. Models can be used for several purposes, including "intelligent interpolation" of patchy infrared or visible data. Where several images are available a few hours apart, with different cloud distribution, this is particularly useful. A model correctly set up for this purpose will allow full use of all the available data, whilst accounting for real temporal changes between the images, e.g. due to tidal advection or heat flux to the atmosphere.

4.4.3 Access to data

Access to data is a critical requirement, in order to exploit the remote sensing information potential to its fullest extent. As a general rule, data acquisition and archiving are best accomplished by dedicated, specialised facilities, usually operated (either directly or under contract) by Space Agencies. As this has been the trend for quite some time, a certain degree of standardisation, extended to end-users interested in having direct data acquisition capabilities, has already been achieved in this field.

Access to data, and long-term continuity of the service, is provided via these central facilities, as inputs to scientific and (quasi) operational applications. However, delays and costs of services that provide data of proven quality, over a number of years, can be a severe limitation to any user. The experience acquired in the Clean Seas context has shown that the vast quantities of data generated by satellite borne sensors, and the scientific complexities inherent in data processing, in particular for optical data, to derive environmental parameters, represent the most serious obstacles to their use. Managing and exploiting the information potential of integrated remotely sensed data requires that substantial efforts be made in the generation of value-added information products, and in their analysis, using specific scientific tools.

The development of specific (quasi) real time data product lines, and of corresponding historical time series - as well as dedicated algorithms and models - in support of marine environmental research and applications, is seen as a pre-requisite to ensure data accessibility. Current data, generated by new sensors, should be -screened, selected, assembled, and archived as quality-controlled sets of derived parameters, at processing levels to be defined. The same tools developed for ingesting new data, as they become available, should be used to update the time series of historical data, possibly coupled to suitable auxiliary data, e.g. atmospheric and meteo-climatic parameters, to enhance the statistical significance of the series or to open new information lines. These activities should also be accompanied by the collection and/or development of data management, processing and analysis tools, to be made available together with the data sets themselves. This should ensure that algorithms and models e.g. site or time specific, are always available for the exploitation of the archived data.

Suitable data access and distribution means, based on electronic publishing techniques, shall also be identified and implemented. Formats (for data archiving, retrieval and distribution) represent a critical issue that should be clarified by competent organisations (e.g. the Committee on Earth Observation Satellites (CEOS) ad-hoc working group). In general, it is envisioned that such implementations should be compliant with the guidelines identified and/or established by current research efforts in this field, e.g. the Centre for Earth Observation (CEO) programme.

4.4.4 Validation of the use of AVHRR for cyanobacteria monitoring

The process of detecting and monitoring the cyanobacterial (blue-green algae) blooms in the Baltic Sea relies on the visual interpretation of NOAA AVHRR channel 1 (visible) imagery. During conditions of strong surface accumulations the cyanobacteria becomes visible also in channel 2 (near-infrared), but not as spatially widespread as in Figure 4-23).

During the intense bloom periods in July-August the interpretation has often been performed by two operators in parallel to validate the results. The differences that occurred was only minor, and in general limited to the exact location of the outer (visible) limit of the bloom. During the 1999 season, when SeaWiFS data was used by one of the operators, the differences were more pronounced due to the much better spectral resolution of the ocean colour sensor.

The Finnish Institute of Marine Research in Helsinki, Finland have supplied in situ data from ‘ships of opportunity’ i.e. automated unattended recordings of fluorescence of chlorophyll a and related parameters (temperature and slinity). The data originate from the ‘Finnjet’ ferry, which used to run between Helsinki and Travemünde, thereby covering a large part of the Baltic Sea on each transect. Based on geographic location, ongoing algae bloom and cloud free conditions during an extended time, two periods were selected for the comparison of AVHRR and in situ data. The first time period was June 4-12 1997, while the second period was August 4-0 the same year. For both occasions two different sections were chosen (Figure 4-24). In spite of the ‘good weather’ during the selected time period, the problem with cloud contamination of the AVHRR imagery still exist.

Figure 4-24. Images showing the Finnjet routes for the selected transect in June (left) and August (right). The green colour denotes a south-bound route.

Instead of comparing a ‘full’ Finnjet transect with the corresponding AVHRR image (actually several images, as the cruise time between Helsinki and Travemünde is approximately 1-1 days), a specific section was selected for analysis.

The first test was to compare fluorescence chlorophyll a data from Finnjet with AVHRR based albedo values from the same day. This comparison did not show any good correlation between the two data types (Figure 4-25). Comparisons from other days at the same or different location indicate more or less similar discrepancies between the in situ data and the AVHRR imagery.

Figure 4-25. Comparison between Chlorphyll a (blue) and AVHRR albedo (orange) on August 9, 1997.

The possible explanation for this is the different measurements being performed. The AVHRR sensor register the intensity at the sea surface from satellite level, while the water being measured by the Finnjet system originate from approximately 2-5 meters depth. This difference in measurement "depth" together with the physio-biological behaviour of the cyanobacteria, e.g. Nodularia spumigena, to accumulate at the sea surface at a late (dying) blooming phase could explain the low correlation for this test. Another factor to consider is the time delay between water intake at the ship hull and the actual fluorescence chlorophyll a determination together with the time stamp. A four minute difference between water intake and actual measurement, at 25 knots, will result in a spatial difference of 3 km.

In the second test which shows a much higher correlation (Figure 4-26), the chlorophyll a values from Finnjet are 5 days earlier than the AVHRR albedo values.

Figure 4-26. Chlorphyll a values from June 7, 1997 (blue), and AVHRR albedo values from June 12 (orange). The two curves are shifted apart for best fit.

This time difference is intended to "compensate" for the behaviour of the algae to float to the surface and thereby becoming visible in the AVHRR imagery at a late bloom stage. To further fit the two curves to each other, they are spatially shifted approximately 5 km. This shift is intended to compensate for the water movement during the 5 days and the possible offset due to the time delay of the chlorophyll a measurements.

The last comparison is between summarised chlorophyll a values against a single AVHRR image (Figure 4-27).

Figure -27. Summarised chlorophyll a values (blue) from June 5,7,9 and 12, 1997; and AVHRR albedo values from June 12.

The chlorophyll a values are from June 5,7,9 and 12, 1999, and are from approximately same geographical points. Some parts of the curves show a good match (with some shift), while other parts tend to show the opposite. The idea of this cumulative comparison was to see the relation between the area with visible surface accumulation in AVHRR imagery and chlorophyll a values the days before.