Seafloor mapping based on multibeam echosounder bathymetry and backscatter data using Object-Based Image Analysis: a case study from the Rewal site, the Southern Baltic

The rapidly evolving seabed mapping technology, related to multibeam echosounders and sidescan sonars, generates large amounts of data, which require time-consuming traditional analyses. For this reason, there is a need for automatic classification of sediment and seabed morphological forms, using the analysis of reflected acoustic signals from the bottom (e.g. Tęgowski & Łubniewski 2002), analysis of images created by the bathymetric model of the seabed, or maps of backscattered intensity of acoustic signals (e.g. Montereale Gavazzi et al. 2016). The analysis of seabed sediment composition, as a branch of marine geology, has recently had more and more in common with hydroacoustic swath measurements. Multibeam echosounder (MBES) surveys allow large areas of the seafloor to be investigated, which is not comparable with traditional sampling technologies. MBES collects two kinds of acoustic data: a bathymetric map and a backscatter intensity image of seafloor areas. While the former generates a precise bathymetric Digital Elevation Model (DEM) that represents seafloor geomorphology, the latter is equal to seafloor acoustic reflectivity, which is characterized by scattering of the acoustic wave by the seabed type (Lurton & Lamarche 2015).

Beyond the MBES bathymetry and backscatter characteristics of the seafloor, other important information on the seafloor is provided by traditional ground-truth samples acquired at point locations. In order to obtain a meaningful map of the seabed, samples often need to be processed at various stages of seafloor mapping. According to Diesing et al. (2016), seafloor mapping usually includes the following steps: pre-processing of a dataset, feature extraction, feature selection, classification, post-classification and evaluation of classifier performance. In this study, pre-processing of data involves processing of hydroacoustic data. The next two steps are described together in “Image processing”. Although this study deals with the broad application of Object-Based Image Analysis (“Image processing”), the use of segmentation-classification algorithms was sufficient to create a final map of seabed sediment composition, and therefore no post-classification of data was necessary. The final step, which involves an accuracy assessment, is described in “Accuracy assessment” section.

The dataset presented in this paper comes from the survey that was conducted by the Maritime Institute in Gdańsk in 2010 on board the r/v Imor ship. Multibeam echosounder measurements and acquisition of ground-truth samples covered 20 km² of the seafloor in the southern part of the Baltic Sea. Analysis of ground-truth samples allows to define four classes of the seabed, three of which were characterized by overlapping backscatter distribution. Despite difficulties, it was possible to process the data and obtain a good classification performance, which was confirmed by an accuracy assessment based on the error matrix.

The hydroacoustic survey was conducted in the area located in the Polish Exclusive Economic Zone (EEZ), in the southern part of the Baltic Sea, north of the coastal village of Rewal. The location and basic parameters of the area are shown in Figure 1.

Figure 1

According to Mojski (1995), the area is covered by recent marine fine- and medium-grained sands, mud and clay from different phases of the Baltic Sea development. General geomorphological features are the result of the glacier impact during the last glaciation, especially the deglaciation and current marine processes that have formed seabed deposits since the Littorina transgression until now (Pieczka 1980; Gudelis & Jemielianow 1982). Geomorphologic structures include accumulation or abrasive plains, as well as shoals with varying thickness of recent marine sand sediments. Clay in abrasive plain areas is often exposed, whereas muddy-clayey sediments from terminal post-glacial basins also occur locally. Accumulation plains are usually filled with muddy-clayey sediments from various phases of the Baltic Sea development, partly covered by sandbanks and sand waves, related to marine transgressions and recent processes that have not yet been fully understood (Pieczka 1980; Gudelis & Jemielianow 1982).

The world’s seafloor has been mapped only in 5–10% with a resolution comparable to onshore research (Wright & Heyman 2008), mainly in areas deeper than 10 m (Montereale-Gavazzi et al. 2016). Although the main device used for hydroacoustic measurements is a multibeam echosounder (MBES), bathymetry is not the most important feature for seabed mapping. The MBES provides additional information about the seafloor reflectivity, or backscatter of the returned acoustic signal, which is crucial for seabed mapping, but its availability is even more limited than that of bathymetric data. There are still no standards for MBES backscatter acquisition and processing, which is why hydroacoustic data collected during one survey is practically incomparable with others (Diesing et al. 2016).

This study deals with the application of a complete methodological approach in order to process the multibeam echosounder backscatter data in one study. The methodology presented in this paper assumes the use of the QINSy-ArcGIS processing method, Object-Based Image Analysis and additional data from ground-truth samples. The in situ approach allows realistic and consistent categorical seafloor mapping (Lurton & Lamarche 2015).

Backscatter grids from the multibeam echosounder are often created using various implementations of the Geocoder engine (Fonseca et al. 2009). This engine makes it possible to apply the Angle Varied Gain (AVG) correction to minimize the ship’s along-track nadir effect and to maximize the contrast of the backscatter intensity image. Among others, the Geocoder engine is implemented in two most common commercial developments: Fledermaus Geocoder Toolbox (FMGT), and CARIS HIPS and SIPS software (Lurton & Lamarche 2015). Both solutions are often used in habitat mapping. For example, FMGT was recently used to create backscatter images in the North Sea (Stephens & Diesing 2014), the Belgian part of the North Sea (Montereale Gavazzi et al. 2017), and the northern China (Li et al. 2017). Raw backscatter data were also processed using CARIS HIPS and SIPS, for example in the areas of Georges Bank in Canada (Brown et al. 2011), Maltese Islands (Micallef et al. 2012), the Tasman Peninsula (Lucieer et al. 2013) and the Lagoon of Venice (Montereale-Gavazzi et al. 2016; Madricardo et al. 2017). For comparison, in this research we applied the new Geocoder engine implementation in the QINSy software. All AVG and mosaicking parameters were tuned and selected carefully. Although the angular dependence of backscatter strength (Parnum 2007; Lurton & Lamarche 2015) could not be completely eliminated, the backscatter mosaic with angular dependence correction, created using this method, gave satisfactory results compared to the backscatter grid without any angular correction. The quality of backscatter mosaic was good enough to perform Object-Based Image Analysis on seabed sediments.

The scale of multiresolution segmentation is a key parameter that may have a major impact on the accuracy of results (Benz et al. 2004). This phenomenon has been observed at a small scale in the Tasman peninsula object-based habitat mapping by comparing the accuracies for 30 and 60 scales (Lucieer et al. 2013). The multiresolution segmentation scale is defined as imperfect, therefore it is necessary to perform many segmentation-classification tunings in order to achieve the desired effect. In order to clearly define a value of the scale parameter, some statistical trials of its estimation were made (Drăguţ et al. 2010; Drăguţ et al. 2014). Nevertheless, multiresolution segmentation results should be visually assessed (Diesing 2016). However, after some unsuccessful applications of the ESP/ESP2 tool in this study, we have decided to use an iterative method of scale selection, as described in “Materials and methods”.

Although it is recommended to acquire and use as many ground-truth samples as possible, statistically more than 50 samples per class (Carlotto 2009), for practical and budget reasons this is not always possible during seabed mapping (Diesing et al. 2016). Therefore, some studies present classification results after using a much smaller but representative number of ground-truth samples (Micallef et al. 2012; Montereale Gavazzi et al. 2017). A similar situation occurs in this study. Nevertheless, attention should be paid to potential sources of errors when many significantly different classes of the seabed occur (Micallef et al. 2012; Diesing et al. 2016).

In marine habitat mapping studies, the distribution of backscatter intensity between ground-truth classes usually shows some scattering (Stephens & Diesing 2014; Montereale Gavazzi et al. 2017). In such a case, the correct use of advanced supervised classifiers should not cause any difficulties and the results may represent a high level of accuracy. The situation is more complex when the backscatter distribution of most classes overlaps. The results of this study show that the proper application of segmentation and classifier algorithms in the Object-Based Image Analysis can contribute to a high accuracy.

The exceptional accuracy of OBIA results in land-cover remote sensing is not always directly reflected in marine habitat mapping (Diesing et al. 2014; Montereale Gavazzi et al. 2016). There are many reasons for this, often related to difficulties in conducting research in a significantly different environment. Nevertheless, the quality of seabed mapping studies is constantly improving due to adaptation of good practices from land cover mapping, greater awareness of image processing techniques, and increased use of geostatistical secondary features (Diesing et al. 2016; Lecours et al. 2016; Li et al. 2016). One of them is rugosity that was selected in this study by the CART embedded feature selection algorithm. As mentioned in section “Image processing”, rugosity is a secondary feature of bathymetry, developed in this study on the basis of the methodology proposed by Sappington et al. (2007). The presented study confirms that even for a dataset with fuzzy boundaries between classes in terms of the distribution of backscatter intensity, the use of other secondary geostatistical layers can lead to a good accuracy assessment. Therefore, it is important to generate such features and use them in the proper selection.

In the past, different decision trees were commonly used in seabed mapping (Dartnell & Gardner 2004; Rattray et al. 2009; Ierodiaconou et al. 2011; Che Hasan et al. 2012a; Che Hasan et al. 2012b; Huang et al. 2012; Stephens & Diesing 2014; Montereale Gavazzi et al. 2016), using both pixel-based and object-based approaches. The term “accuracy” is typically associated with the predictive performance of a map and a reference ground-truth dataset (Foody 2002). A literature review of 20 publications on marine habitat mapping conducted by Diesing et al. (2016) shows that a validation subset of ground-truth data was used in 77% of the analyzed studies. The overall accuracy and the Kappa index were most frequently used. No error matrix was found in 2/3 of the studies. Che Hasan et al. (2012a,b) used the Quick, Unbiased and Efficient Statistical Tree (QUEST), a decision tree method based on image segmentation, resulting in the overall accuracy of 80.2% and the Kappa statistic of 67%. The object-based CART method was investigated by Montereale Gavazzi et al. (2016), but the results were not described, because they were less promising compared to the K-NN (K-Nearest Neighbor) classifier. The remaining studies represent a pixel-based application of decision tree classifiers. For example, Dartnell & Gardner (2004) show results with the overall accuracy of 72%. The Kappa index was not calculated in this study. Ierodiaconou et al. (2001) evaluated the pixel-based type of the QUEST decision tree with the overall accuracy of 80% and the Kappa statistic of 75%. The classification tree (CT) was used by Stephens & Diesing (2014) with the same overall accuracy and a lower Kappa coefficient (48%, but the highest of all tested classifiers). Huang et al. (2012) used the Boosted Decision Tree (BDT), a slightly different pixel-based method to calculate seabed sediment parameters. The error matrix and the overall or Kappa measures were not included in this case. Depending on the study, decision trees produced different results. The high accuracy obtained in this research (overall accuracy = 87.5%, Kappa = 81%) confirms the high usefulness of the Object-Based Image Analysis and the CART classifier. The range of the Kappa index is equal to [–∞, 1], and the result below 0.8 is interpreted as good, while close to 1 as very good.

The paper presents the general workflow of backscatter mosaic processing with the angular dependence correction, using common processing software for a multibeam echosounder dataset. For the first time in the peer-reviewed literature, we have described the QINSy-ArcGIS method for multibeam backscatter data processing, including the application of the AVG correction of the Geocoder engine in the QINSy software. Depending on the size of the dataset, the QINSy-ArcGIS method can be time-consuming, but in some cases it may be the only way to produce or reprocess a good quality backscatter mosaic based on limited or few-years old backscatter MBES data. Once created, it can be successfully fitted to other MBES datasets, reducing the total processing time.

The results of this research confirm the high usefulness of the methods of Object-Based Image Analysis and decision trees, including Classification and Regression Trees (CART). Tuned and properly applied Object-Based Image Analysis can be used as a powerful tool to create a meaningful map of seabed sediments, even for ground-truth classes with overlapping backscatter distribution. In this case, another layer beyond backscatter can improve the classification accuracy. The use of OBIA, including CART and the feature selection tool, produces good accuracy assessment results (overall accuracy = 87.5%, Kappa = 81%). Therefore, we propose to add the CART algorithm to other classification methods used in future comparative studies.