Coastal ecosystems are essential because they support high levels of biodiversity and primary production, but their complexity and high spatial and temporal variability make their study particularly challenging. Seagrasses are extremely important marine angiosperms (flowering plants) with a worldwide distribution. Seagrass meadows are among the most productive ecosystems in the world, which help protect the shoreline from soil erosion, serve as a refuge area for other species, and absorb carbon from the atmosphere [1
]. Thus, seagrasses are essential, and their preservation in a sustainable manner needs the appropriate management tools. In this sense, satellite remote sensing is a cost-effective solution that has many advantages, compared to traditional techniques, like airborne photography with photo-interpretation or in-situ measurements (binomic maps from oceanographic ships). This way, satellite remote sensing is becoming a fundamental technology for the monitoring of benthic habitats (e.g., seagrass meadows) in shallow waters, as it provides periodic and synoptic data at different spatial scales and spectral resolutions [3
Seafloor mapping using satellite remote sensing is a complex and challenging task, as optical bands have limited water penetration capability and the best channels to reach the seafloor (shorter wavelengths) suffer from higher atmospheric distortion. Hence, the signal recorded at the sensor level coming from the seabed is very low, even in clear waters [4
]. Towards the goal of mapping benthic habitats at high spatial resolution and achieving a reasonable accuracy, the use of hyperspectral (HS) imagery can be considered as an alternative to multispectral (MS) data. Unfortunately, high spatial hyperspectral sensors onboard satellites are not yet available and, in consequence, high-resolution data from airborne or drone HS sensors are the only options to collect HS data to map complex benthic habitats environments.
To map the seafloor, the use of high-resolution remote sensing is promising but requires the application of different geometric and radiometric corrections. Specifically, the removal of the atmospheric absorption and scattering and the sunglint effect over the sea surface are essential preprocessing steps. In addition, the water column disturbance can be corrected; however, it is a very complex issue in coastal areas due to the variability of the scattering and absorption in the water column, the bottom type, and the water depth [6
Regarding the removal of the atmospheric effects, correction approaches can be basically grouped into physical radiative transfer models and empirical methods exclusively considering information obtained from the image scene itself [7
]. Many scene-based empirical approaches have been developed to remove atmospheric effects from multispectral and hyperspectral imaging data [8
]. Concerning the physical models, they are more advanced, complex and based on simulations of the conditions of the atmosphere from its physical-chemical characteristics and the day and time of acquisition of the image. At the present time, there are a number of model-based correction algorithms, for example MODerate resolution atmospheric TRANsmission (MODTRAN), Atmosphere CORrection Now (ACRON), Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH), High-accuracy Atmospheric Correction for Hyperspectral Data (HATCH), Atmospheric and Topographic CORrection (ATCOR), or Second Simulation of a Satellite Signal in the Solar Spectrum (6S) [12
]. Some of these algorithms include more advanced features, such as spectral smoothing, topographic correction, and adjacency effect correction.
On the other hand, the removal of sunglint is necessary for the reliable retrieval of bathymetry and seafloor mapping in shallow-water environments. Deglinting techniques have been developed for low-resolution open waters and also for high-resolution coastal applications [15
]. In general, algorithms use the near-infrared (NIR) channel to eliminate sunglint assuming that water reflectivity in the NIR band is negligible [16
]. This assumption is usually correct, except when turbidity is high, or the seabed reflectance is important, which can occur in very shallow areas [6
Concerning the water column correction, Lyzenga [17
] proposed the depth invariant index (DII), an image-based method to decrease the water column attenuation effect. This correction technique has been applied in previous works, due to its simplicity, with different degrees of success [18
]. On the other hand, in the last decades, some radiative transfer models have been proposed, but they are more complex, and the difficulty of accurately measuring some in-situ water parameters can limit their applicability [22
Once preprocessing algorithms have been applied, classification techniques can be used to generate the seabed maps. Classification is one of the most active areas of research in the field of remotely sensed image processing. For example, the classification of hyperspectral imagery is a challenging task because of the imbalance among the high dimensionality of the data and the limited amount of available training samples, as well as the implicit spectral redundancy. For this reason, specific approaches have been developed, like random forests, support vector machines (SVMs), deep learning or logistic regressions [26
]. Unmixing techniques have also attracted the attention of the hyperspectral community. Unmixing algorithms separate the pixel spectra into a collection of constituent pure spectral signatures, named endmembers, and the corresponding set of fractional abundances, representing the percentage of each endmember that is present in the pixel [27
Recent research to create seabed maps using remote sensing imagery has been mainly devoted to map coral reefs [28
] or seagrass meadows [3
]. Commonly, these studies address very shallow, clear and calm waters, and very dense vegetal species (i.e., Posidonia oceanica
). As a continuation of our preliminary study [14
], in this work, hyperspectral and multispectral imagery have been used to compare the benefits of each type of data to map the seafloor in a complex coastal area where submerged green aquatic vegetation meadows have low density, are relatively located at considerable depths (5 to 20 meters), where the sea surface is usually not completely calm due to persistent local surface winds, and, consequently, where very few bands reach an acceptable signal-to-noise ratio. Hence, a thorough analysis has been performed to obtain a robust methodology to produce accurate benthic habitat maps. To achieve this goal, different corrections, object-oriented, and pixel-based classification approaches have been considered, and diverse feature extraction strategies have also been tested. In summary, contributions are presented regarding the best correction techniques, feature extraction methods and classification approaches in such a challenging scenario. Moreover, a comparative assessment of the benefits of satellite multispectral and airborne hyperspectral imagery is included to map the seafloor in complex coastal zones.
3. Results and Discussion
After the correction of each dataset (see Figure 4
), three supervised classifiers were applied to different combinations of input data. All the analysis was performed for HS and MS imagery (AHS and WV-2, respectively) at the complex area of Maspalomas.
As seafloor reflectivity is very weak, and following the steps of Figure 4
, precise preprocessing algorithms were applied to correct limitations in the sensor calibration, solar illumination geometry, viewing effects, as well as the atmospheric, sunglint, and water column disturbances. In this sense, geometric, radiometric and atmospheric corrections were performed. As specified, 6S was selected to model the atmosphere and to remove the absorption and scattering effects in the multispectral image [14
], and ATCOR4 for the hyperspectral data [49
]. This selection took into account the results of a previous validation campaign comparing real sea surface reflectance recorded by a field spectroradiometer (ADS Fieldspec 3) and the reflectance estimated for WV-2 data using different models. Next, deglinting algorithms were applied to eliminate the solar glint and whitecaps over both datasets. Finally, the seafloor albedo was generated applying the radiative transfer model described in Section 2.4.1
. As shown in Figure 5
, for a small area, the improvement is considerable, especially for the AHS data, as some areas were severely affected by de-sunglint.
A preliminary analysis was performed to find the most suitable corrected imagery to address the classification problem. Specifically, images obtained after the different pre-processing steps were assessed to identify the more reliable data source for the mapping production. The following thematic classes were considered: sand (yellow), rocks (brown) and Cymodocea
(green). Using the information from the ship transects and sampling sites (Figure 2
c), sets of training and validation regions were generated including regions of each class at five-meter step depths from 5 to 20 m. Approximately, 3000 and 6000 pixels per class were selected for the training and test ROIs, respectively. The class pair separability (Jeffries-Matusita distance [52
]) in the bands ranges from 1.218 to 1.693 for WV-2 and between 1.802 and 1.985 for AHS. These values corroborate a better discrimination capability of HS data as more spectral richness is available.
presents the results of applying the three supervised classifiers to the data after the atmospheric, sunglint, and water column corrections. The same independent training and validation regions of interest were used in all the experiments. As expected, the airborne hyperspectral imagery allows a better classification than the satellite multispectral data (92.01% with respect to 88.66%). It can be appreciated that the best overall accuracy was achieved after the deglinting step. The water column removal did not improve the seafloor mapping, even after applying a complex radiative model. Even providing adjusted water IOPs and bathymetry values, the modeling of the background albedo by linear mixing of benthic classes in this complex area does not seem adequate for the subsequent classification. The very low reflectivity of the coastal bottom, which usually contributes less than 1% of the radiation observed by the sensor, produces errors in the adjustment of the abundances of the modeled pure benthic elements. Clearly, the model considered has to be further improved. As indicated, the water column modelling in coastal areas is complex and depends on the water quality parameters, as well as the bathymetry and the type of seabed. For this reason, this preprocessing is not always considered and some studies demonstrate that better results are not always achieved [18
]. Finally, regarding the classification algorithm, SVM is the most appropriate approach for AHS but Maximum Likelihood works better with WV-2.
shows examples of seafloor maps generated for the AHS sensor, with SVM, and for the WV-2 data using the ML classifier. Comparing the results with the reference benthic map (Figure 2
d) and the available video records from the ship transects, higher accuracy can be noted for AHS and using the imagery after the atmospheric and sunglint correction (middle row). Excessive amount of submerged vegetation is identified for WV-2 and some rocks incorrectly appear on the right side when these pixels should be labeled as vegetation.
To improve the previous seabed cartography, a detailed feature extraction and classification assessment was only performed using the preprocessed data after the atmospheric and sunglint correction stages.
As stated in Section 2.4.2
, to improve the benthic maps, additional information was obtained using feature extraction techniques. In particular, PCA, ICA and MNF were applied to the corrected spectral bands. In the analysis, the classifier performance was assessed including the complete new set of components after these transforms and, in addition, the best components were also tested discarding noisy bands. Figure 7
shows the first bands of each transform, as well as the original spectral bands as a reference (the remaining bands were not displayed as they are too noisy). Regarding the spectral channels, we can appreciate that only shorter wavelengths (first bands) can reach the seafloor and, in consequence, even dealing with hyperspectral data only a few channels are really valuable to map benthic habitats up to a depth of 20 m. On the other hand, PCA, ICA and MNF provide useful information in the first four components. The true color image and false color composites using the first three components are also included, and it is possible to check the worse behavior of ICA and the noise removal effect of MNF.
Additional textural parameters and abundance maps after unmixing were also inputted to the classifiers as auxiliary information.
summarizes the AHS and WV-2 accuracy results of each classifier for the following input combinations:
Spectral bands after atmospheric and sunglint corrections.
Components after the application of three-dimensionality reduction techniques (PCA, ICA, and MNF). The complete dataset and a reduced number of bands or components were both tested.
Abundance maps of each class after the application of linear unmixing techniques.
Texture information (mean and variance) extracted from the first PCA/MNF component.
Pixel-based classification was applied to the previous options and, finally, object-based classification was applied to the spectral bands.
With respect to the sensors, we can appreciate that AHS provides better accuracy than WV-2, as expected, mainly due to the availability of additional bands and a better radiometric resolution. Specifically, a major improvement is attained for SVM (mean accuracy increase of 9.5%) than for ML (4.7% average increase).
Concerning the classification algorithms, SAM did not work properly because, even being more insensitive to variations of the bathymetry, classes are spectrally overlapped, and only very few bands are useful due to the water column attenuation. SVM is the algorithm achieving the best accuracy, but the simpler and faster ML demonstrates good performance and, in many cases, better than SVM (average results in Table 3
confirm it). Actually, Figure 8
presents the comparative performance of both classifiers, and it can be appreciated that ML is more robust, providing more stable results regardless of the input information used or the number of bands considered. Specifically, the standard deviation (averaged for AHS and WV-2) of the overall accuracy for the different combinations is 2.9% for ML and 6.4% for SVM.
PCA and MNF perform much better than ICA, but the improvement is, in general, negligible with respect to the original bands. Also, the reduction of the number of bands/components to avoid the Hughes’ phenomenon is basically not increasing the classification accuracy except for ML and the hyperspectral data. The number of training pixels for each class is high enough (3000), and that could be a possible explanation.
The application of unmixing techniques before the classification did not improve the accuracy due to the small number of bands actually available. It can be appreciated the degraded performance of SVM when only the three abundances are considered in the classification scheme.
Finally, texture information is a feature that can be included in the final methodology as precision values in some circumstances increase the performance. Specifically, the improvement is more evident for SVM and using the texture information provided by the first component of PCA. For ML, texture generally does not provide a better accuracy of the benthic map.
It is important to highlight that results obtained by the object-based classification techniques (OBIA) are not always the best. Basically, OBIA only provides superior performance than pixel-based techniques for the SVM algorithm. However, results are quite dependent on the type of segmentation considered.
In general, the overall accuracies for ML and SVM are high as few classes are considered and the validation pixels chosen to numerically assess accuracy were selected in clear and central locations of each seabed type. In any case, the relative results between the different classifiers and input combinations displayed in Table 3
includes an example of the AHS and WV-2 segmentation for a specific area. AHS provides more detailed information and, in consequence, the number of objects increases.
compares the best pixel-based seafloor maps generated by ML and SVM for the AHS image. A majority filter of 5 × 5 window size was applied to remove the salt and pepper effect. Both maps are very accurate, but ML overestimates vegetation (green) in some specific areas, while SVM the rocks (brown) in others. Finally, Figure 11
shows the best maps for each sensor obtained using the object-based classification with the SVM algorithm. Results are similar and, in general, match the available eco-cartographic map included in Figure 2
d, except for the western side of the area. In any case, as indicated, this map was just considered a coarse reference and really ship transects T1 and T2 in Figure 2
c demonstrate the existence of vegetation meadows in that area, in agreement with AHS and WV-2 maps. It is also important to highlight that a vulnerable and complex ecosystem was studied where the density of submerged green aquatic vegetation beds is quite low and, therefore, there is a considerable mixture of sand and plant contributions in each pixel of the image.
These methodologies will be shortly applied to generate precise benthic maps of natural protected ecosystems in other vulnerable coastal ecosystems. In addition, these will be applied to hyperspectral imagery recorded from drone platforms with the goal of discriminating between the different vegetation species.
A comprehensive analysis was performed to identify the best input dataset and to obtain a robust classification methodology to generate accurate benthic habitat maps. The assessment considered pixel-based and object-oriented classification methods in shallow waters using hyperspectral and multispectral data.
A vulnerable and complex coastal ecosystem was selected where the submerged green aquatic vegetation meadows to be classified are located at depths between 5 and 20 meters and have low density, implying the availability of very few spectral channels with information and a considerable mixing of spectral contributions in each image pixel.
Appropriate and improved atmospheric and sunglint correction techniques were applied to the HS and MS data. Next, a water radiative transfer model was also considered to remove the water column disturbances and to generate the seafloor albedo maps. A preliminary analysis was performed to identify the most suitable preprocessed imagery to be used for seabed classification. Three different supervised classifiers (maximum likelihood, support vector machines, and spectral angle mapper) were tested.
A detailed analysis of different feature extraction methods was performed with the goal to increase the discrimination capability of the classifiers. To our knowledge, the effect of three rotation transforms to generate benthic maps was assessed for the first time. Texture parameters were, as well, added to check whether spatial and context information improve classifications. Finally, the inclusion of abundance maps for each cover, obtained by the application of linear unmixing algorithms, was also considered but, given the small number of spectral bands actually reaching the seafloor, results were not fully satisfactory. The best results were produced by SVM and the OBIA approach. However, to generate benthic habitat maps, the simple ML has shown an excellent performance and superior stability and robustness than SVM (average overall accuracies over 3% and 7% for AHS and WV-2 data, respectively).
In summary, a robust methodology was identified, including the best correction techniques, feature extraction methods, and classification approaches, and it was successfully applied to multispectral and hyperspectral data in a complex coastal zone.