Remotely sensing the German Wadden Sea—a new approach to address national and international environmental legislation

  • Gabriele Müller
  • Kerstin Stelzer
  • Susan Smollich
  • Martin Gade
  • Winny Adolph
  • Sabrina Melchionna
  • Linnea Kemme
  • Jasmin Geißler
  • Gerald Millat
  • Hans-Christian Reimers
  • Jörn Kohlus
  • Kai Eskildsen
Article

DOI: 10.1007/s10661-016-5591-x

Cite this article as:
Müller, G., Stelzer, K., Smollich, S. et al. Environ Monit Assess (2016) 188: 595. doi:10.1007/s10661-016-5591-x

Abstract

The Wadden Sea along the North Sea coasts of Denmark, Germany, and the Netherlands is the largest unbroken system of intertidal sand and mud flats in the world. Its habitats are highly productive and harbour high standing stocks and densities of benthic species, well adapted to the demanding environmental conditions. Therefore, the Wadden Sea is one of the most important areas for migratory birds in the world and thus protected by national and international legislation, which amongst others requires extensive monitoring. Due to the inaccessibility of major areas of the Wadden Sea, a classification approach based on optical and radar remote sensing has been developed to support environmental monitoring programmes. In this study, the general classification framework as well as two specific monitoring cases, mussel beds and seagrass meadows, are presented. The classification of mussel beds profits highly from inclusion of radar data due to their rough surface and achieves agreements of up to 79 % with areal data from the regular monitoring programme. Classification of seagrass meadows reaches even higher agreements with monitoring data (up to 100 %) and furthermore captures seagrass densities as low as 10 %. The main classification results are information on area and location of individual habitats. These are needed to fulfil environmental legislation requirements. One of the major advantages of this approach is the large areal coverage with individual satellite images, allowing simultaneous assessment of both accessible and inaccessible areas and thus providing a more complete overall picture.

Keywords

Satellite remote sensing Monitoring Seagrass Mussel beds Environmental EU directives Wadden Sea 

Introduction

The Wadden Sea, a coastal wetland of 450 km length along the North Sea coasts of Denmark, Germany, and the Netherlands, is one of the largest wetlands in the world. It forms the transition zone between the North Sea and the mainland and is characterized by large changes in water coverage due to the semi-diurnal tides. The Wadden Sea is a highly dynamic area with habitats comprising tidal channels, sandbars, mudflats, and saltmarshes. Its physical characteristics provide the basis for a high primary productivity, which in turn leads to a high biodiversity. Benthic invertebrates such as bivalves, gastropods, crustaceans, and worms form a rich food source for higher trophic species such as fishes and birds. The Wadden Sea is a key site for millions of migratory birds and its ecosystem sustain wildlife populations well beyond its borders.

In Germany, the Schleswig-Holstein Wadden Sea National Park was established by law in 1985, the corresponding national parks in Lower Saxony and Hamburg following in 1986 and 1990, respectively. The national park laws together with the Federal Nature Conservation act provide the basis for monitoring operations in these areas. The Trilateral Cooperation of Germany, Denmark, and the Netherlands protects and conserves the Wadden Sea as an ecological entity through common policies and management since 1978. In 1997, the harmonized Trilateral Monitoring and Assessment Programme (TMAP) was implemented, providing a common basis for the monitoring and assessment of the Wadden Sea ecosystem, management and policy (TMAG—Trilateral Monitoring and Assessment Group 2009). Consequently, the resulting standardization of methods permits an assessment of the whole Wadden Sea as well as the comparison of different regions.

Due to its high ecological value and importance, the Wadden Sea was stated as Special Protection Area according to the Birds Directive (BD; 79/409/EEC), and major parts have been designated as Special Areas of Conservation according to the NATURA 2000 network within the framework of the Habitats Directive (HD; 92/43/EEC). Large areas are also contained within the River Basin Districts of the Water Framework Directive (WFD; 2000/60/EC) and were declared as Wetlands of International Importance according to the Ramsar Convention. More recently, the Dutch, German, and Danish areas of the Wadden Sea were recognized as UNESCO Natural World Heritage site, underpinning their ecological value. The number of environmental directives affecting the Wadden Sea increased in the last decade (lastly the Marine Strategy Framework Directive (MSFD); 2008/56/EC) similar to the demands concerning monitoring and reporting obligations. In Germany, the results from TMAP are used to address many of these European monitoring requirements. However, some of the currently implemented monitoring schemes suffer from a highly reduced accessibility of the area due to tidal changes in water coverage, which only permit visual or personal access to the tidal flats during a small time period around low tide. In addition, muddy sediments or the occurrence of tidal channels may hamper field surveys of larger areas or the access from land to certain areas.

Because of these logistical difficulties, investigations on the applicability of satellite remote sensing for assessing intertidal surfaces started in the early 1980s using data from the Multispectral Scanner (MS) and Thematic Mapper (TM) aboard Landsat (Pröber 1981; Dennert-Möller 1982; Folving 1984). While many studies focused on sediment classification and distribution (Bartholdy and Folving 1986; Allewijn 1992; Yates et al. 1993; Kokke 1995), water and vegetation cover as well as topography also became of interest to the scientific community (Doerffer and Murphy 1989; Wang and Koopmanns 1995). Advances in satellite technology not only allowed increases in spatial and/or spectral resolution (e.g. Landsat-8, RapidEye, or WorldView), but also concurrently promoted the development of image classification methods. For intertidal areas, these include maximum-likelihood classification (Yates et al. 1993; Kokke 1995), unsupervised classification and principle component analyses (Doerffer and Murphy 1989; Kleeberg 1990; HIMOM 2005), application of linear spectral unmixing followed by decision tree classification (Rainey et al. 2003; Stelzer and Brockmann 2006; Brockmann and Stelzer 2008), hierarchical classification approaches (Jung and Ehlers 2014), and multiple regression between spectral bands and surface characteristics (Rainey et al. 2000; Smith and Fuller 2001; Deronde et al. 2006; Adam et al. 2011; Ibrahim and Monbaliu 2011). In recent years, the combination of optical and SAR sensors for classification of intertidal flats has been investigated (van der Wal and Herman 2007; Gade et al. 2008; Gade and Stelzer 2010).

Concurrent with the technology-pushed advances in satellite remote sensing, there has been an increasing awareness of the opportunities this technology provides to fulfil the requirements from nature protection managers for environmental information. Particularly with regard to the increasing amount of monitoring obligations from national and international legislation, the cooperation of nature protection managers and remote sensing experts is of utmost importance in turning remote sensing technology into a useful tool for monitoring and management. Kennedy et al. (2009) provide guidelines for cooperation between nature protection managers and remote sensing scientists in order to meet monitoring objectives. Fraser et al. (2009) developed a framework for monitoring land cover change and ecological integrity of six Canadian national parks using Landsat TM and ETM+ time series data. In Greece, remote sensing data have been used for long-term monitoring as well as updating of habitat maps of wetlands, which were fully or partly designated as Ramsar sites (Alexandridis et al. 2009; Bortels et al. 2011). Together with technological advances, the open data policy for imagery from sensors such as Landsat and the Sentinels also promote the inclusion of remote sensing techniques into monitoring applications.

While airborne remote sensing techniques, i.e. aerial mapping and imagery, have for long been included in operational monitoring programmes for seagrass and mussels, satellite remote sensing of coastal areas has become a potential alternative only recently, due to advances in sensor technology. In this paper, a framework for a synergistic classification of the German Wadden Sea using various satellite remote sensing data is presented. The framework has been developed as a collaboration of nature protection managers from the national park administrations and other state agencies together with experts in optical and SAR remote sensing, within three research projects: OFEW (2005–2007), DeMarine-Environment TP4 (2008–2011), and DeMarine-2 SAMOWatt (2012–2015). The framework is used to provide area-wide information on selected monitoring parameters, e.g. mussel beds, seagrass meadows, and sediment classes. This information is needed to meet monitoring requirements from, amongst others, the WFD, HD, and MSFD. The applicability of the resulting products for monitoring purposes in the German Wadden Sea is discussed.

Methods

Mussel beds and seagrass meadows constitute two important habitats in the Wadden Sea. Together with bare sediments, they represent different surface types for which remote sensing products have been developed. While seagrass and sediments are classified based on optical satellite data only, the identification of mussel beds, due to their inherent surface roughness, highly benefits from the inclusion of radar data into the classification process. In this section, a description of the general classification framework is given, followed by a detailed representation of the classification of mussel beds and seagrass meadows.

General classification framework

The general framework for the classification of intertidal flats consists of five steps (Fig. 1). Different satellite sensors have been included in our studies. Optical image data were obtained from Landsat (-5, -7, and -8), RapidEye, and SPOT-4 sensors, whereas synthetic aperture radar (SAR) data were acquired from the TerraSAR-X (TSX) sensor.
Fig. 1

Schematic representation of the general classification framework used for the development of products for addressing monitoring requirements in the Wadden Sea

Step 1 of the classification framework includes the selection and preprocessing of the optical and SAR remote sensing data. For both data types, subsetting and masking of land areas are needed. For optical image data, a simple atmospheric correction using the dark pixel subtraction and geometric correction was carried out. Thus, the bases for the following steps are the surface reflectances in the dedicated wavelengths of a sensor. Radiometric calibration, co-registration, speckle filtering, and geometric correction were performed for SAR data. The retrieved normalized radar backscatter cross section is used as input for the following step.

In the second step of the classification framework, the feature extraction takes place. The features are derived by linear spectral unmixing or consist of band ratios or first derivatives of the reflectance spectrum for optical data (Brockmann and Stelzer 2008). Multi-temporal statistical parameters and polarization coefficients are extracted from the SAR data (Gade et al. 2014; Gade and Melchionna 2016). The resulting features act as input parameters for the classification.

The third step, the classification itself, is based on a decision tree which recursively partitions the data set into smaller subdivisions based on a range of tests at each branch of the tree. Instead of the commonly used statistical approach, this framework uses knowledge-based tests on one or more features for creating the subdivisions according to available spectral and/or textural information on the different classes. This makes the classification process independent of the actual number of pixels belonging to a certain class. Different decision trees have been developed for vegetation and mussel classification, sediment types, seagrass density classification, and water coverage.

The fourth step in the framework is the validation of the resulting classifications using data from long-term monitoring programmes as well as in situ data collected during the different projects (ground truth data). The validation uses areal and positional comparisons as well as confusion matrices for selected surfaces. The in situ data were gathered using sampling protocols especially designed for this purpose and addressing remote sensing relevant parameters (e.g. sediment type and colour, vegetation cover, and roughness scales). The fieldwork was timed as close as possible to satellite acquisitions in order to ensure an optimal comparability of in situ data and satellite images.

The product generation (step 5) includes the extraction of the relevant classes from the classification process and the creation of the final map products with respective metadata.

Dedicated classification of mussel beds and seagrass meadows

Study area and surface types

For this study, four different test sites in the German Wadden Sea were selected (Fig. 2). The first test site is located south of the island of Norderney in the Lower Saxony Wadden Sea. The second test site encompasses the major part of the North Frisian Wadden Sea. The two remaining test sites, Amrum and Langeneß, are located northwest and west of the island of Amrum. The intertidal flats in these areas are characterized by the occurrence of mussel beds, which differ in size as well as proportion of blue mussel (Mytilus edulis) vs. Pacific oyster (Crassostrea gigas). Furthermore, the largest seagrass meadows of the whole Wadden Sea are located in the North Frisian Wadden Sea.
Fig. 2

Location of the four test sites in the Wadden Sea of Lower Saxony and Schleswig-Holstein: Norderney, North Frisia, Amrum, and Langeneß

Blue mussels and Pacific oysters may form extensive mussel beds in the Wadden Sea, covering tens of hectares. Their ecological importance mainly lies in the filtering of sea water and in the formation of habitats for a large number of smaller species (Dame et al. 1991; Dankers and Zuidema 1995; Nehls et al. 2009). Typically, mussel beds are found on elevated intertidal flats. They may contain areas of bare sediments and water puddles of different sizes (Fig. 3). The colours of the mussels range from dark blue for the blue mussels to a yellowish white for the Pacific oyster. Mussel beds are often covered by the perennial brown algae (Fucus) which can be detected by remote sensing with relatively high accuracy due to the strong vegetation signal in the optical data. Fucus gives a strong signal in NIR the whole year round so that this information can also be used in the non-vegetation period to aid the identification of mussel beds.
Fig. 3

Mussel bed showing the typical structure with elevated patches of mussels (partly covered with Fucus) interspersed with water puddles (left) and a high-density (>60 %) seagrass meadow (right)

Mussel shell fields are spectrally similar but much brighter in the visible part of the spectrum and are thus collected in a separate class in the classification process. Compared to seagrass and bare sediments, the surface structure of mussel beds is very rough, making the inclusion of radar data for their identification very meaningful.

Seagrass meadows provide food, habitat, and nursery areas for a wide range of invertebrate and vertebrate species and are thus important for the Wadden Sea ecosystem (van der Graf et al. 2009). The two seagrass species Zostera marina and Zostera noltii have growing seasons from May to October, with the meadows reaching their maximum extent in August and September. The largest seagrass meadows on the German coast can be found in Schleswig-Holstein between the islands of Föhr and Pellworm.

Remote sensing data

The optical sensors Landsat, SPOT, and RapidEye are suitable for the identification of mussel beds and seagrass meadows (Brockmann and Stelzer 2008). However, while Landsat’s 30-m pixel size is at the lower limit of useable resolution, RapidEye’s resolution of 6.5 m together with the RedEdge band is very suitable and even allows the detection of smaller patches. All sensors have bands in the visible and near-infrared (NIR) spectrum, which is necessary for the detection of mussels and seagrass. Bands in the short-wave infrared are very suitable for correcting the measured signal for water coverage, but are not available with all sensors. Within this paper, we show the results of SPOT-4, RapidEye, and Landsat-7 and Landsat-8 data to which the classification scheme has been applied (Table 1).
Table 1

Acquisition dates and captured regions of optical and radar images used in the present study

Date

Sensor

Region

27-07-2008

SPOT-4

North Frisia

26-07-2009

SPOT-4

North Frisia

30-04-2012

RapidEye

Amrum

23-06-2012

TSX

Norderney

21-07-2012

TSX

Norderney

26-07-2012

Landsat-7

North Frisia

01-10-2012

TSX

Norderney

05-06-2013

TSX

Amrum

06-06-2013

TSX

Amrum

22-06-2013

TSX

Amrum

05-08-2013

TSX

Amrum

10-08-2013

TSX

Norderney

15-08-2013

Landsat-8

North Frisia

19-08-2013

TSX

Amrum

21-08-2013

TSX

Norderney

30-09-2013

Landsat-8

Norderney

19-04-2014

Landsat-8

North Frisia

17-07-2014

Landsat-8

North Frisia

It has already been shown that SAR sensors are useful for the identification of mussel beds (Choe et al. 2012; Gade et al. 2014; Gade and Melchionna 2016). Multi-temporal statistics benefit from the fact that the radar signal from mussel beds is strong and rather independent of radar look direction, incidence angle, and weather conditions. The signals of bare sediments, seagrass meadows, and water puddles, on the other hand, depend highly on these technical and environmental parameters and are therefore very variable. A comparison of mean and standard deviation from multi-temporal statistics may thus provide indicators for mussel beds. The SAR data used for this study comprise five images of the Amrum area from 2013 and five images of the Norderney area from 2012 and 2013 (Table 1).

Classification procedure

After preprocessing and feature extraction, the input parameters are entered into the classification scheme (Fig. 1). The decision tree for mussel identification is also used for identification of vegetation and provides several classes: one class for mussels, one class for mussels covered with water, one class for shell detritus, three different vegetation density classes, one class for bare sediment, and one class for water. Decisions leading to a mussel class combine the optical characteristics of high reflection in the NIR with low reflection in the red part of the visible spectrum, namely the Normalized Difference Vegetation Index (NDVI), and the gain between red and NIR reflectances. Besides the strong reflectance in the NIR, Fucus is almost black and thus has no strong green peak in the visible part of the spectrum. This fact is used by calculating the distance of the green band from the baseline between the blue and red bands from the optical sensors. In addition, statistics from multiple SAR acquisitions, namely the mean and standard deviation of the normalized radar backscatter cross section, are used as features in the decision tree for a better separation of mussel beds from vegetated areas.

The strong vegetation signal derived from seagrass is captured by the NDVI which can be categorized into different density classes (10–20, 20–60, >60 %) by applying thresholds to the NDVI. Even a class 5–10 % is included, mainly classifying diatoms on the sediment flats. These respective thresholds have been calibrated with seagrass density information from ground surveys performed in parallel to satellite overflights. For synergistic classifications, one optical image is combined with the multi-temporal statistics derived from multiple SAR images. Finally, the classes of interest are extracted from the classification and converted into polygon vector data.

Validation

Validation is performed using in situ data from mussel and seagrass monitoring programmes from the National Park authorities in Schleswig-Holstein and Lower Saxony. Depending on their overall extent, a large part or even all mussel beds are mapped during regular ground surveys. The remaining mussel beds are identified from aerial images. The data are available as shapefiles and provide, amongst others, information on location and areal extent of individual mussel beds.

Seagrass is monitored by aerial mapping covering the whole Wadden Sea each year. In addition, the extensive and dense seagrass meadows in Schleswig-Holstein are mapped by ground surveys covering one sixth of the entire area each year, providing a whole picture with additional information every 6 years. Both monitoring methods provide shapefiles with two different density classes, 20–60 and >60 %. Since the growth rate of seagrass depends on different environmental factors, e.g. light, nutrients, temperature, and algal overgrowth (Lee et al. 2007), a good temporal alignment of image acquisition and monitoring date is essential for a meaningful comparison.

The classification data are loaded into a geographical information system together with the data from the monitoring programmes. Visual inspection of data layers provides an estimate of location accuracy. Areal comparisons are performed at the scale of individual mussel beds or seagrass meadows as well as at larger scales, i.e. a complete test site such as North Frisia. The reliability of classifications can be assessed using multi-year comparisons, but natural changes have to be considered in this approach.

Results

This chapter presents the resulting classifications and the comparison of the classifications with ground truth data for mussel beds and seagrass meadows.

Overall classification of mussel beds and seagrass meadows

Figure 4 shows the synergistic classification of mussel beds and seagrass density classes for two Landsat images acquired at different dates, each combined with the mean backscatter at horizontal polarization derived from five TSX images acquired in 2013. The upper Landsat scene was acquired in August 2013 and shows large areas of seagrass indicated by dark red areas in the false-colour representation (upper left) and in shades of green in the classification map (upper right). The second example (lower row) is based on the Landsat image acquired in April 2014. This date is outside the vegetation period and thus the classification shows almost no vegetation.
Fig. 4

Results from the synergistic classification of the Amrum area based on Landsat-8 images acquired on August 15, 2013, (upper left) and on April 19, 2014, (lower left) combined with the mean backscatter derived from five TSX images from 2013 (middle images). The results of the synergistic classifications are shown to the right

In the Landsat-8 scene from August 2013, the water level is very high, but many of the water-covered mussel beds still appear in the classification. This is primarily due to the results from the radar backscatter acquired during lower water levels which are combined with the weak vegetation signal from the optical data.

Synergistic classification of mussel beds

The combination of optical and SAR data has a strong impact on the identification of mussel beds. The improvements in the classification of mussel beds are demonstrated for the Norderney test site using a Landsat-8 image from September 2013 and statistics from five TSX images from 2012 to 2013 (Fig. 5). If only optical image data is used in the classification process, significant parts of the mussel beds will be classified as vegetation due to the Fucus cover (left). The SAR data add information on the surface structure and thus improve the distinction between Fucus-covered mussel beds and vegetation on bare sediment. The synergistic classification therefore detects larger parts of the mussel beds and shows a better agreement with the data from the mussel monitoring programme (right). The areal extent of the mussel beds shown in Fig. 5, identified by optical and synergistic classification, together with the percent agreement with the monitoring data is shown in Table 2. The inclusion of SAR data into the classification scheme raises the agreement with the monitoring data from 53 to 79 %.
Fig. 5

Comparison of optical (left) and synergistic (right) classifications of mussel beds in the Norderney test site. A Landsat-8 image acquired on September 30, 2013, was used in both cases. Mean backscatter was derived from five TSX images acquired in 2012 and 2013. The black lines indicate mussel bed data obtained during the 2013 monitoring campaign

Table 2

Comparison of areal extent of mussel beds shown in Fig. 5 at the Norderney test site using different methods and percent agreement with monitoring data

Method

Area (ha)

Percent agreement with monitoring data

Optical classification

37.8

53

Synergistic classification

57.1

79

Monitoring

72.3

 
For a quantitative assessment of the synergistic classification, eight mussel beds from the Amrum and Norderney test sites were selected (Fig. 6). The mussel beds have been selected based on different conditions, which are expected to affect the classification accuracy. The two lower rows in Fig. 6 permit a visual comparison of the classification results with the monitoring data. The quantitative assessment of the areal agreement between classification and monitoring data is shown in Fig. 7.
Fig. 6

Synergistic classification of individual mussel beds from the test sites Amrum (Landsat-8 image from April 19, 2014, together with SAR data from 2013) and Norderney (Landsat-8 image from September 30, 2013, together with SAR data from 2012/2013). The location of the test sites is shown in the upper left image and the location of the individual mussel beds within the test sites are shown in the upper middle and right images. The middle and lower row show the classified mussel beds together with the monitoring data for the Amrum and Norderney test sites respectively

Fig. 7

Comparison of the areal agreement between the synergistic classification and the monitoring data for selected mussel beds from the Amrum and Norderney test site (see Fig. 6 for the location of the mussel beds)

The best agreement was found for mussel beds located in high intertidal areas exposed to short periods of submersion during the tidal cycle (Fig. 6a-c,e,f). With respect to mussel beds located in lower intertidal areas (Fig. 6h) or in intertidal creeks (Fig. 6d,g ), the classification tends to underestimate their size. These mussel beds are covered by water for a longer time, thus reducing the chances to be captured by the satellite sensor. In some cases, however, even mussel beds located in tidal creeks may be classified with reasonable accuracy (Fig. 6g). These examples demonstrate the influence that environmental conditions and location effects have on classification accuracy.

If mussel beds are submerged by water or covered by a thin mud layer, the classification is less accurate or even fails. Fucus is a reliable indicator for mussel beds because it needs hard substrate to attach to and therefore is often associated with mussel beds. Under these circumstances, its vegetation signal, particularly in the non-vegetation period, can be used to distinguish between mussel beds covered with a thin mud layer and bare sediments. However, the actual mussel beds are not seen by the sensor and the classification can be less accurate. A thin mud layer or the occurrence of Fucus does not basically affect the surface roughness of mussel beds. Therefore, the use of SAR data and a corresponding synergistic classification will allow the detection of mussel beds under these circumstances.

In order to be applicable as a monitoring method, the classification scheme needs to be transferable to different sensors and the results have to be repeatable over time. For one mussel bed at the Amrum test site (Fig. 6c), the classification has been performed for four different years with optical images from different sensors and in two cases as synergistic classification with optical and TSX data (Fig. 8). This mussel bed is quite stable and therefore well suited for demonstrating the usability of the classification approach for routine monitoring purposes. Apart from two cases, the classification results agree well with those from the monitoring campaigns (Fig. 9).
Fig. 8

Classifications of a mussel bed in the Amrum test site (mussel bed c in Fig. 6) based on either pure optical data (a, b, c, and f) or on optical and radar data (d and e). The different vegetation density classes have been combined. SYN synergistic classification, L8 Landsat-8. The black lines show the monitoring data for the respective year

Fig. 9

Comparison of mussel bed areas derived from classification (dark grey) and monitoring campaigns (light grey) for the same dates as shown in Fig. 8

Larger differences were found for July 2009 (SPOT-4) and for July 2014 (Landsat-8). In both cases, the classification is based on optical data only, since no SAR data were available for the respective time periods. For July 2009, the area derived from monitoring data is significantly larger than in later years, and in the area of the monitored mussel bed, the classification yielded a mixture of vegetation and mussels. This artefact often occurs in classifications based on optical data only. For July 2014, the classification resulted in a large area of shells in the northern part, which were not included in the monitoring data and thus the classification overestimated the size of the mussel bed.

Optical classification of seagrass

The extent and density of the seagrass meadows around Langeneß can be clearly seen in the false-colour infrared image as red areas (Fig. 10a). While exposed sediments show up in grey and light-blue colours, water-covered sediments and open water appear in blue colours. The classification (Fig. 10b) yielded three different density classes: 10–20, 20–60, and >60 %. The two latter classes correspond to those that are used in the regular monitoring programme. The classification quality was visually assessed using information from aerial mapping (Fig. 10c) and ground surveys (Fig. 10d). Since the ground surveys only cover one sixth of the complete area per year, a temporal match with a suitable satellite overflight is very seldom. Therefore, different years had to be compared in this study. It appears that all three methods provide roughly similar results in terms of areal extent and density for the seagrass meadows around Langeneß, even though the time difference precludes a quantitative comparison. However, the level of detail that the classification provides, with respect to delineation of seagrass areas and the capture of tidal creeks as well as bare sediments within the meadows, is much higher.
Fig. 10

Classification and reference data for seagrass at the test site Langeneß: a Landsat-8 false-colour infrared image acquired on August 15, 2013; b classification of Landsat-8 image mentioned in a; c seagrass monitoring data derived from aerial mapping (August 2013); d seagrass monitoring data derived from ground surveys (2010)

A quantitative comparison of all seagrass meadows in the North Frisian Wadden Sea between 2008 and 2014 is shown in Fig. 11. For this comparison, the 10–20 % density class obtained in the classification has been included in the 20–60 % density class. This was done because this class provides a clear vegetation signal which distinguishes it from the surrounding bare sediments. In addition, one inherent problem of the aerial mapping method is an uncertainty in determining the lower density limit, which therefore may not lie at exactly 20 %. Furthermore, it has to be taken into account that seagrass density can change very rapidly during the growing season and a time difference of 1–3 weeks may significantly alter the results. The comparison shows that aerial mapping tends to produce higher total areal estimates than the classification. Considering the methods and the degree of detail each of them captures, this tendency was expected.
Fig. 11

Area of seagrass meadows in North Frisia (period 2008–2014) determined using optical image classification (SPOT-4 in 2008 and 2009, Landsat-8 in 2013 and 2014) and aerial mapping data. Ground truth data was totalled for the 6-year period and included as horizontal lines for comparison

Discussion

Suitability of satellite remote sensing for mussel monitoring

This study showed that the use of satellite remote sensing has great potential to significantly contribute to the monitoring of mussel beds in the Wadden Sea. In particular, this applies to the large spatial coverage of individual images, which allow a complete coverage of either the Lower Saxonian or North Frisian Wadden Sea at individual points in time. This is highly important for the rapid detection and assessment of changes, e.g. after storms or winters with drift ice. The large spatial coverage also permits the detection of emerging and/or hitherto unknown mussel beds in areas that are particular inaccessible and therefore not generally surveyed during field work. The analysis of the satellite images provides standardized products, for which the detection is independent of the observer/analyst. Products from different seasons or years can therefore easily be compared and used for change detection analyses. Using local satellite imagery, the techniques presented herein can be transferred to other regions of the Wadden Sea and even other intertidal flat areas without any difficulty. A major advantage of this method is the transferability of reflectance measurements into features. This allows data from different optical sensors to be used as input for the classification scheme. Finally, the technique, specifically the synergistic classification, has also been shown to be applicable if the mussel beds are covered with a few centimetres of water during the overflight of the optical sensor. This further extends the time frame around low tide for acquiring suitable optical imagery, thus increasing the possibility of obtaining products within specified time frames.

Even though slight water coverage does not markedly influence mussel bed detection, high tide or cloud cover often prevents acquisition of suitable imagery at specific times. Together with the relatively long revisit times of sensors such as Landsat, this means that product delivery at predefined points in time currently cannot be guaranteed. The occurrence of Fucus with its strong vegetation signal is a reliable indicator for mussel beds particularly during the non-vegetation period. The signal might be confused with similar signals from dense seagrass beds or green algae mats during summer and autumn, making the differentiation more difficult at these times. In the present study, the inclusion of SAR data markedly improved the results. In contrast, a reduced detection capability for both optical and radar techniques was found due to the deposition of layers of mud, which affects both the mussel beds’ natural colour and surface roughness. This problem, however, is not specific to satellite remote sensing but also applies to the interpretation of aerial imagery, which is used in the current monitoring programmes.

In general, while products for mussel bed identification based on satellite remote sensing techniques cannot yet occur in a timely reliable manner, they may still provide valuable information which can be integrated into current monitoring programmes. The large spatial coverage of satellite images can be used to generate complete maps of the Wadden Sea area. These can then be used as basis for the planning of field surveys, thus optimizing the use of resources. The visual interpretation of aerial imagery may also benefit from such maps since it relies heavily on surface colour which, in many areas, makes it difficult to distinguish between dark mussels and wet sand or mud. Furthermore, information on emerging and/or disappearing mussel beds not visited during field surveys can be obtained on a regular basis. With respect to the TMAP and the European environmental legislation, these maps may thus help to address two of the main monitoring parameters, namely the location and extent of individual mussel beds in the Wadden Sea.

Suitability of satellite remote sensing for seagrass monitoring

Our results showed that seagrass meadows can be classified based on optical satellite data with a high degree of detail. It is possible to distinguish up to five different density classes, which is beyond the two density classes required by the environmental directives and the three classes presented in this study. Due to the highly variable growth rate, the seagrass density as well as meadow extent may change rapidly over the course of just a few days. This makes the comparison with and validation against data from the current monitoring programme difficult. On the one hand, the ground survey is very accurate, but suffers from the fact that only one sixth of the area is surveyed each year. In addition, the annual survey is not done on a single day but rather over a time period of a few weeks, during which some seagrass meadows are very likely to have changed considerably. On the other hand, aerial mapping provides an overview of the whole area, but suffers from rather coarse delineation of seagrass meadows as well as positional displacements. Therefore, seagrass distribution maps based on satellite remote sensing techniques might constitute a valuable supplement to current monitoring schemes, provided that suitable imagery can be acquired. The latter currently represents the bottleneck in remote sensing of seagrass meadows in the Wadden Sea.

While the seagrass classification can distinguish several different density classes, it cannot currently distinguish between the two seagrass species Z. marina and Z. noltii. Furthermore, a dense layer of diatoms as well as mats of green algae may also provide strong vegetation signals, which may be confused with seagrass meadows. However, due to the long-term monitoring of seagrass, all suitable locations for seagrass meadows in the Wadden Sea are known and this information can therefore be used to eliminate wrongly classified seagrass meadows.

Current and future applicability of remote sensing products for monitoring in the Wadden Sea

Apart from mussel beds and seagrass meadows, the distribution and dynamics of sediments have also been investigated in the OFEW and DeMarine projects. The accuracy of the classifications was validated by in situ field data and expert knowledge. Parameters such as percentage of sand, mud, and hard substrate have been extracted for the different water bodies in order to address monitoring requirements of the WFD in the year 2009 (MLUR 2010).

Even though the classification products for mussel beds and seagrass meadows have not yet been used to address environmental legislation directly, they have already proven their usefulness in different ways. The large spatial coverage of individual images allows acquisition of parameters at individual points in time, which is of particular importance for seagrass as described above. The large coverage implies that an overview of e.g. North Frisia or Norderney can be obtained at individual dates and used as basis for ground surveys, thus optimizing both financial and human resources. In addition, information can be obtained on remote or inaccessible areas. Acquisition of several images per year would permit change detection within comparatively short time frames. Necessary management measures could therefore be implemented without large time lags. The standardized classification algorithms allow a detection of parameters independent of user experience, which makes the results more comparable over time compared to ground surveys (Hearn et al. 2011; Stevens et al. 2004). Furthermore, by using regional imagery, the classification framework can be applied to all areas of the Wadden Sea and can thus provide comparable information for the three neighbouring countries. The use of single images to address several parameters may contribute to an integrated and cost-effective monitoring, the latter being an important factor for environmental agencies with generally limited financial resources.

Currently, the use of remote sensing products for monitoring purposes in the Wadden Sea is mainly limited by the unpredictability of obtaining suitable imagery, both optical and radar, due to water coverage or weather. This is exemplified by the lack of suitable images for seagrass classification in the years 2010, 2011, and 2012 (Fig. 11). Quantitative comparisons with monitoring data did not always produce high agreement values for both mussel beds and seagrass meadows. Particularly for seagrass meadows, this might have been caused by different acquisition dates of satellite and field data. Furthermore, each monitoring method has its specific advantages and disadvantages, which have to be taken into account when comparing and assessing the classification results against the monitoring results. Aerial mapping of seagrass, for example, provides an overview over North Frisian seagrass meadows within a short time frame. In this respect, it is comparable to the classification of a single satellite image. However, the amount of detail available in the satellite-based classification together with its locational accuracy exceeds the quality of the aerial mapping data by far.

Considering the advantages and disadvantages of using the remote sensing products described here, it becomes clear that remote sensing products cannot substitute current monitoring programmes. They can, however, offer significant additional value and may perhaps be able to substitute parts of the monitoring programmes, e.g. aerial mapping of seagrass, in the future. In order to improve the current remote sensing products, the first tests are already being performed with data from Sentinel-1 and Sentinel-2, which have shorter revisit times and, in case of Sentinel-2, higher spatial and improved spectral resolution than most satellite data used in this study. For the next 3 to 5 years, current Wadden Sea monitoring programmes will run in parallel with remote sensing classification in order to build a solid basis for future decision-making.

Conclusions

We have developed a new approach for the routine monitoring of intertidal areas that is based on a combined use of remotely sensed optical and radar data. The use of satellite remote sensing techniques aids in tackling one of the main difficulties for monitoring programmes in Wadden Sea areas, the accessibility of vast areas during only short periods at low tide. In this respect, one of the main advantages of using satellite remote sensing data is their large areal coverage, which allows a simultaneous assessment of large areas at individual points in time.

The classification scheme proposed herein consists of five main steps: (1) preprocessing of the input data, (2) feature extraction, (3) classification, (4) validation, and (5) product generation. The specific data and methods used in each step vary depending on the final product.

Our results show that both mussel beds and seagrass meadows can be detected with reasonable accuracy. The classification of mussel beds benefits from the inclusion of radar data, because those beds often have significantly rougher surfaces, thereby allowing their detection even when they are covered by a thin mud layer or by a few centimetres of water. In addition, emerging and/or disappearing mussel beds in inaccessible areas can be monitored without the need of field surveys. Furthermore, up to five different density classes for seagrass meadows can be distinguished, which is better than current monitoring requirements. The quality of the seagrass classification, with respect to location accuracy and density information, exceeds by far that based on aerial mapping methods.

Currently, the main bottleneck lies in the availability of suitable imagery due to weather conditions (optical data) or water coverage (both optical and radar data), which often hampers the generation of classification products. Those are, however, of particular importance for the routine monitoring of habitats such as seagrass meadows, for which their maximum extent and density has to be determined annually. Nonetheless, and despite these limitations, the resulting products already offer substantial added value, since they provide area-wide information that can be used to optimize field campaigns.

During the next few years, the new classification scheme based on remote sensing data will run in parallel to current Wadden Sea monitoring programmes, thereby providing a solid basis for decision-making. Furthermore, new satellites such as the recently launched European Sentinels may well be able to further increase the quality and reliability of products for the monitoring of the Wadden Sea.

Acknowledgments

We thank H. Büttger and H. Farke as well as all colleagues and volunteers who assisted with the data collection during field campaigns. F. Leverenz and F. Werner helped in the processing of TSX data. Furthermore, we like to thank M. Nyenhuis and colleagues from the German Aerospace Center (DLR) for constructive support during the SAMOWatt project. We also thank the Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research on Sylt, BioConsult SH in Husum, and Syltair on Sylt, as well the National Park Authorities in Lower Saxony and Schleswig-Holstein for providing monitoring data and/or logistical support during data collection. TerraSAR-X data were provided by DLR under contract OCE0994. Data from Landsat-7 and Landsat-8, RapidEye, and SPOT-4 were obtained from the USGS, RESA, and the ESA Third Party Mission.

Funding was received from the German Ministry of Economy (BMWi) for the projects DeMarine SAMOWatt (contract numbers 50EE1112, 50EE1115, 50EE1117) and DeMarine-Environment TP4 (contract numbers 50EE0830, 50EE0816, 50EE0817).

Compliance with ethical standards

Conflict of interest

The authors declare that they have no conflict of interest.

Copyright information

© Springer International Publishing Switzerland 2016

Authors and Affiliations

  • Gabriele Müller
    • 1
  • Kerstin Stelzer
    • 2
  • Susan Smollich
    • 2
  • Martin Gade
    • 3
  • Winny Adolph
    • 4
  • Sabrina Melchionna
    • 3
  • Linnea Kemme
    • 3
  • Jasmin Geißler
    • 1
  • Gerald Millat
    • 4
  • Hans-Christian Reimers
    • 5
  • Jörn Kohlus
    • 1
  • Kai Eskildsen
    • 1
  1. 1.Schleswig-Holstein Agency for Coastal Defence, National Park and Marine Conservation, National Park AuthorityTönningGermany
  2. 2.Brockmann Consult GmbHGeesthachtGermany
  3. 3.Institute of OceanographyUniversity of HamburgHamburgGermany
  4. 4.National Park Authority Wadden Sea of Lower SaxonyWilhelmshavenGermany
  5. 5.State Agency for Agriculture, Environment and Rural AreasFlintbekGermany

Personalised recommendations