1. Introduction
Terrestrial remote sensing has a comparatively long history since the 1970s and benefits from the availability of land surface imagery at various spatial and spectral resolutions, supplemented by ancillary data such as environmental data and digital terrain models. This enabled the development of sophisticated classification methods and algorithms that extract critical information for environmental management from remote sensing imagery [
1,
2]. Conversely, seabed automatic classification efforts are more recent [
3,
4,
5] and mostly based on acoustic mapping technologies, since the application of electromagnetic methods has severe limitations due to the rapid attenuation of light in water.
The development of new technologies and classification tools is in great demand from the industry, government agencies and scientific institutions. The increasing interest and necessity for seabed habitat mapping in marine spatial planning initiatives has forced the industry of sonar makers, software developers and scientific research institutions to invest in new technology capable of producing seabed classification over regional areas [
6]. Fisheries, oil and gas exploration, offshore engineering, as well as shallow and deep sea mining are examples of industries that are requiring seabed classification to support environmental risk assessments and spatial planning [
6,
7,
8].
Side scan sonars (SSS) and multibeam echosounders (MBES) are among the most widely applied acoustic techniques for seabed mapping. Both systems have high sampling rates, which enables seabed mapping at a broad spatial scale, producing high resolution data sets in a relatively short period of time [
9,
10,
11,
12].
A series of classification models have been developed to reduce the subjectivity and automate the interpretation of geophysical, sedimentological and imaging data sets obtained from the seabed. Supervised and unsupervised classifications have improved the seabed classification [
13,
14] with the potential to also compile statistical analyses of those datasets.
The study presented here attempts to automatically classify and map submerged reefs. The identification and mapping of reefs through automatic classification from swath-mapping systems has great potential and applicability due to the global importance of these complex habitats, regarding planning and management of marine protected areas (MPA) and fisheries resources [
15,
16,
17].
The analyses were carried out based on a dataset collected in the Abrolhos Continental Shelf (Eastern Brazilian Continental Margin), where diverse reef morphologies, including pinnacles, reef banks and rhodoliths, have been observed [
18,
19]. In order to automatically classify the distinct reef types, two supervised classification techniques were applied, taking into account that each of them has a different type of data input and distinct architecture. One technique is a morphometric classification tool (Benthic Terrain Modeler, BTM), with focus on the spatial analyses of the bathymetry derivative properties. The second technique is an object-based image analysis (OBIA, using eCognition software) applied only to the backscatter data from the SSS.
We have also compared the different methods using an assessment of map accuracy. To achieve this, the classification maps obtained from BTM and eCognition were compared against a reference dataset. In this context, the aim of this contribution is to analyze the potential benefits and limitations of the distinct techniques in the recognition and automatic classification of reefs through the complementary use of MBES or SSS or both data sets.
4. Discussion
We have presented the results of two automated methods (BTM and OBIA) for mapping reefs in a test site situated in the Abrolhos outer reef arc. The accuracy of these methods in detecting reef against a background of non-reef was estimated by comparing them with a reference map that was obtained by detailed manual interpretation of SSS backscatter and MBES bathymetry data. In the following, we will discuss the obtained results qualitatively and quantitatively.
The estimation of map accuracy requires a reference dataset that can be acquired in different ways, but needs to fulfill at least one of the following two conditions: (1) The reference source data has to be of higher quality than the data used for map classification, or (2) if using the same source data for both map and reference classification, the process to create the reference data needs to be more accurate than the process used to generate the map classification to be evaluated [
27]. Here, we use the same datasets, but both SSS and MBES data were used to create the reference map, while the two methods utilised either one or the other. Also, we are satisfied that the process of creating the reference map is more accurate due to the following reasons: (1) There exists substantial expertise from previous studies [
18,
19] on which the interpretation is based, (2) the reference map is the consensus of the interpretation of three skilled experts, and (3) reefs are conspicuous features that are easily recognizable in the acoustic data. We acknowledge that the reference map is not “the truth”, but under the given circumstances we are satisfied that it is close to reality and can be used as a reference against which the other methods are assessed.
The quantitative analysis of agreement with the baseline map showed that both methods, despite being very different in terms of input data and methodology, gave comparable results regarding map agreement and error. The map agreement of 67% (OBIA) and 69% (BTM) is acceptable, although lower than what has been reported in some recent studies that mapped reef [
28,
29,
30]. These differences might be related to various factors, e.g., shallow warm-water coral reefs can be mapped with satellite or drone-based optical sensors, which give more spectral information than a one-band SSS backscatter dataset. However, this also indicates that there is room for future improvements and the potential to transfer mapping approaches from the optical remote sensing community [
4]. The use of the accuracy statistics proposed by [
26] gives some interesting insights: for both methods, disagreement is predominantly associated with quantity, i.e., the proportions of predicted reef and non-reef differ from the reference. This is also reflected in the high errors of commission for reef presence and is due to the over-estimation of reef area by both approaches, mainly attributable to reef pinnacles (see
Figure 8a,b). Conversely, allocation disagreement is relatively low (around 10%) for both methods, indicating that the overall patterns and trends are mapped correctly.
Automated OBIA classification based on SSS backscatter intensity did not show seabed differences among inter-reef areas. Conversely, this was possible with the morphometric BTM classification. In these regions, the image characteristics (such as texture) did not respond to the differentiation presented by the BTM. Knowing that distinct morphometric characteristics are not necessarily reflected through SSS backscatter intensity, this technique’s complementarity was essential to achieve a better result, even with this limitation in the image classification.
Regarding reef structure classification (
Figure 8d), the results by both automatic techniques were quantitatively similar considering the area covered by reef and disregarding reef sub-type (for example pinnacle or low relief bank). The BTM classification resulted in a coverage area of reef structures (classes: “Inter Reef Structures”, “Pinnacles/IRRB”, “Pinnacles/IRRB-artifacts”, “Edges of LRRB”, “LRRB”) of 6.85 km
2, while the OBIA classification gave an area value of 7.61 km
2 (classes: “Pinnacles”, “LRRB” and “IRRB”).
Reef classes, such as pinnacles, IRRB and LRRB, were not completly differentiated by both techniques. The BTM approach did not differentiate the isolated Pinnacles from IRRB (the value for the sum of the classes “Inter Reef Structures”, “Pinnacles/IRRB-Artifacts” and “Pinnacles/IRRB“ was 3.04 km2), while the OBIA method gave values of 1.57 km2 and 1.75 km2 for Pinnacles” and IRRB, respectively. The OBIA classification provides a better differentiation of these reef features due to their textural characteristics and form being more easily distinguished.
The LRRBs were spatially overestimated by both automatic techniques. The key factor for the overestimation with BTM was the BPI, since some regions presented this index in a similar way to the morphometric characteristics of those banks. When considering the image segmentation in the OBIA process, the selected scale parameter probably contributed to the overestimation of LRBB. In other words, since one of the objectives was to recognize isolated pinnacles, a large number of segments were generated in the image and this made some of these segments more susceptible to small changes in image characteristics (shape, color, and texture), and therefore the classification became particularly vulnerable to these slight differences.
While under the given circumstances, i.e., mapping of reefs, which exhibit specific morphometric and textural characteristics against a relatively flat and featureless background (non-reef areas) manual mapping can be an appropriate approach, it must be recognized that this is nevertheless a slow and painstaking process, especially when carried out by three analysts as it was done here. It is therefore unlikely that such a mapping approach is a viable option for larger seafloor areas. In this study, we mapped approximately 3 km
2 of reef, while the reef area in the wider Abrolhos Shelf was estimated to 8844 km
2 [
18]. By contrast, automatic techniques can deal more efficiently with larger datasets. It might still be necessary to adapt classification dictionaries and rule-set parameters, but the general workflows could still be applied to other (larger) datasets. A viable process might be to develop reference maps for selected test areas, which are characteristic for the wider area. Then, automatic methods can be trialled and improved, with quantitative assessments of agreement and error as tools to decide on the most promising method.
Alternatively, the results of different automatic approaches could be combined to yield an ensemble map [
31]. Such an approach is most efficient with complementary methods, which was the case in this study. By combining the results of the two methods, it was possible to create a new map, which does only include areas classified as reef by both methods. The new map was derived by multiplying BTM with OBIA. For this purpose, we consider reef presence as a value of 1, and absence as a value of 0. This final analysis (full details in
Supplementary information S3) shows the over-estimation of reef extent is reduced (Agreement = 75.3%), and the argument of the complementarity of the methods is strengthened.
Future directions for comparing automatic mapping with a manual reference map should consider the combined use of bathymetric and backscatter dataset. Here, we used two different approaches considering the data availability and computer processing limitations, but OBIA could also be tested using both datasets and a higher resolution grid, e.g., 1-m grid cell. Further investigations in this reseacrh should consider the biological description in order to test if the reef morphology could be used as a proxy for reef benthic community changes along water depth. Also, the use of a higher resolution multibeam sonar could provide a better detail mapping of the Abrolhos reefs. The pinnacles in Abrolhos present a specific morphology that changes with depth [
32]. Shallower reefs (<10 m deep) have wide tops, forming a mushroon-like morphology. Deepward, reef tops decrease in width and can form a single column at depths around 25–30 m. This was observed using scuba diving, but the resolution of the multibeam sonar used in this study was not able to map these detailed changes in pinnacle morphology.