The multibeam echosounder (MBES) bathymetry utilised in this study was obtained from the INFOMAR hydrographic dataset, which is freely accessible on the INFOMAR website (https://www.infomar.ie) ( Figure 1 ). Bathymetric data at 10 m resolution were downloaded and processed using ESRI ArcMap v 10.6. Firstly, fine holes in the dataset were filled with the mean of the surrounding 5x5 pixel neighbours. A general median filter (5x5 rectangle) was applied to remove ‘salt-and-pepper’ imperfections and fine artefacts before re-gridding using a nearest neighbour algorithm. For the purpose of this large-scale mapping, a resolution of 25 m/pixel was deemed a good compromise between morphological detail, partial suppression of acquisition artefacts in the INFOMAR dataset (especially at the outer beam) and computing speed. Bathymetry derivatives were calculated using ArcMap built-in algorithms or with the help of the Benthic Terrain Modeller (BTM) toolbox version 3.0 (Walbridge et al., 2018). The derivatives created include three bathymetric position indexes, a vector ruggedness measurement, two aspect functions (eastness and northness) and three types of hillshades ( Table 1 ). The aspect functions rasters were smoothed using a Gaussian filter (5x5 rectangle) to simplify the signal and reduce salt-and-pepper effects.

For the purposes of FCNN model training, the bathymetry and derivatives were normalised to a double precision value between 0 and 1 based on the minimum and maximum value recorded or calculated in the case for derivative layers.

To train the weakly supervised convolutional neural network, we had to define a dataset from which the model could learn the relationship between bathymetry and derivative data and the landforms present. So firstly, a suitable classification system had to be chosen. The classification system adopted is derived from the Mareano-INFOMAR-Maremap-Geoscience Australia (MIM-GA) two-part marine geomorphology scheme, a standardised seabed mapping glossary aimed to enable more consistent seabed classifications (Dove et al., 2016; Dove et al., 2020). This framework independently describes seabed features according to their observed physical structure (Morphology), and the more subjective interpretation of their origin and evolution (Geomorphology). The separation between physical structure and genesis aligned well with the scope of the machine learning-based mapping of this study, where classes were defined based upon the textural characteristics of the surface rather than apparent geological nature or proper geomorphological definitions. This mapping approach was chosen both because of the general lack of geological ground-truthing for novel marine datasets, and for the exploratory nature of the exercise. In general, defined classes describe archetypal seabed textures which, in various combinations, form seabed landforms. The basic distinction between sediment and rock landforms was nonetheless retained (see Table 2 ), and the INFOMAR sediment grab dataset (https://www.infomar.ie/maps/interactive-maps/seabed-and-sediment) was consulted at the labelling stage ( Figure 1 ). Three principles for classification were adopted (following the advice in Barrett et al., 2022): (a) classes had to be representative of the diversity of seabed morphologies encountered on the Irish continental shelf. This is a “completeness” rule; FCNNs classify pixels in maximum-likelihood fashion, therefore it is essential to fully capture the problem domain as the FCNNs cannot create a new class, or leave a space blank, if an unknown type of seabed is encountered. (b) The classification sets were kept simple and short, as a comprehensive, lengthy list of classes would potentially create difficulties in the training process and especially create more subjective inconsistencies during the labelling work carried out by the expert, and (c) last, but most importantly, classes had to be distinct so that their differences could be confidently isolated visually by the human mapper. This step is critical as the labelling stage may introduce subjectivity and inconsistencies in class delineation that can affect the capabilities of the networks. Therefore, care was taken to semantically define each class, making sure that delineation could be performed with a high level of confidence notwithstanding the limited geological knowledge.

Fully Convolutional Neural Networks (Long et al., 2014) are an extension of traditional CNN architectures (Krizhevsky et al., 2017) adapted for semantic segmentation. CNNs comprise a series of layers that process lower layer inputs through repeating convolution and pooling operations followed by a final classification layer. Each convolution and pooling layer transform the input image, or in this case bathymetric data, into higher level abstracted representations. FCNNs can be broken down into two networks: an encoder and a decoder network. The encoder network is identical to a CNN, except the final classification layer is removed. The decoder network applies transposed convolutions in order to up-sample feature maps back to the original input size, and each decoding stage combines corresponding feature maps created by the encoder network. The final classification layer utilizes 1 by 1 convolution kernels (Lin et al., 2014) to transform the original bathymetric data and derivatives source into a set of dense probabilities using a softmax transfer function. Network weights and biases are adjusted through gradient descent by minimizing the loss function between network outputs and the ground truth pixel labels.

The overall architecture of the FCNN used in this study ( Figure 3 ) is based on a U-Net (Ronneberger et al., 2015) and the encoder networks are VGG-13 and ResNet50 (Simonyan and Zisserman, 2015; He et al., 2017). Residual learning using ResNet encoders has proven to surpass very deep neural networks such as VGG, but for completeness in results we experimented with both encoder networks. The decoder network applies a transposed 2 by 2 convolution for a learnt up-sample track and concatenates feature maps from the encoding network at appropriate resolutions followed by a final 3 by 3 convolution. The final 1 by 1 convolution condenses feature maps to have the same number of channels as the total number of classes in the dataset ( Figure 3 ).

Semi-supervision is the process of incorporating unlabelled image samples for the optimisation of deep neural networks. This branch of deep learning methods is more applicable when unlabelled data are readily available, while labelled instances are often hard, expensive, and time-consuming to collect. Semi-supervised methods can be capable of building better classifiers that compensate for the lack of labelled training data and therefore present a cost-effective solution to label acquisition. In this study, where semantic segmentation was achieved with a pixel classifier, the masks that were used to label pixels did not cover entire 256x256 pseudo-images and therefore every pseudo-image had pixels that were left unlabelled. This condition allowed for unsupervised loss terms to be added into the optimisation process and thus for semi-supervision to be implemented. The supervised loss term is calculated by processing a mini batch of images X ∈ ℝ B × C × H × W and corresponding segmentation maps Y ∈ ℝ B × C × H × W , where B, C, H and W are batch size, number of input channels, height and width. The network produces per-pixel logits Y ¯ ∈ ℝ B × K × H × W - where K is the number of target classes. The softmax transfer function (1) converts network scores into probabilities by normalizing all K scores for each pixel to sum to one:

Where, x ∈ Ω ; Ω ⊆ ℤ 2 is a pixel location and P k ( x ) is the probability for the k t h channel at pixel location x. The negative log-likelihood loss is calculated between segmentation maps and network probabilities:

For each image, the supervised loss is the sum of all losses for each pixel using Equation (2) and averaged according to the number of labelled pixels in Y. Full details on the use of semi-supervision can be found in Hobley et al. (2021). The training parameters and convergence of FCNNs was analysed by testing multiple settings for learning rate and batch size, and assessing computed confusion matrices over several consecutive runs of the algorithm. This ensured that a fair range of different convergence approaches was evaluated. Furthermore, for the semi-supervised approach, several different loss weights were experimented to tune for the unsupervised loss term. The best performing networks were trained for 300 epochs with a batch-size of 12 using AdamW optimiser with a learning rate set to 0.001. With regards to the semi-supervised approach, the unsupervised loss was scaled down by a factor of 10 and the confidence threshold for teacher prediction was set to 0.97. All FCNNs were implemented and trained using Pytorch version 10.2, the code is freely available on our GitHub repository: https://github.com/BrandonHobley/geomorph_deep.

The quantitative metrics of interest to evaluate a classification algorithm are precision and recall (Equations (3) and (4)). These metrics are adequate to test classification algorithms over different datasets as well as their capability to detect false positives and false negatives. Precision and recall are metrics that can show how a classifier performs for each specific class, where precision measures the ability of the model to identify only the relevant instances, while recall measures the ability to detect correctly the occurrence of a class of interest. For instance, in a dataset with 17 confirmed landforms and 121 false landforms, an algorithm that detects every case as false would have an accuracy of 87%, but at the same time it would have an extremely poor recall of 13%. The F1-score (Equation (5)) is the harmonic mean of recall and precision, giving a suitable generalised single figure of merit to convey the performance of a classifier.

The quantitative evaluation metrics listed above are valid if the dataset is labelled, which in our study covers a small subset of the total surface mapped. Therefore, these results can give an indication of a particular classifier performance, but visual inspection is still required to fully grasp the capabilities of FCNNs for bathymetric data.

For this study, 40 different FCNN model runs were carried out, and total mean of their performances are presented in Table 3 . Precision, recall and F1 scores for each class and model are instead given in the Supplementary Material .

The first set of results (models #1 to #12) shows the initial tests carried out on the two different encoders (VGG13 and ResNet50) comparing their efficacy and assessing the utility of semi-supervision and some preliminary combinations of input layers. Overall, the scores show that neither VGG13 nor ResNet50 outperforms the other, although ResNet50 produces slightly better scores at the second decimal point, with increases between 0.01 and 0.05 (e.g. compare models #6 and #7 or #10 and #11).

The use of semi supervision does not improve significantly nor consistently the results, contributing to positive or negative fluctuations. For example, ResNet50 model #11 acquires 0.053 points in Precision compared to non-supervised #12 ( Figure 4 ), with no change in Recall. ResNet50 model #8 gains only 0.004 points in Precision and loses 0.004 points in Recall compared to unsupervised model #7. VGG13 models seem instead to suffer more the application of semi-supervision, leading to higher loss in scores (e.g. compare model outputs #5 and #6).

The best model results were achieved using the complete set of input layers, with ResNet50 models #11 and #12 giving F1 scores of 0.724 and 0.688 respectively. Evaluation metrics are supported by the visual assessment of the resulting thematic maps, where models #11 and #12 show the most visually pleasing results ( Figure 4 ). Nonetheless, high scores were also obtained limiting the input to a combination of bathymetry, BPIs, VRM and Aspect functions (models #3 and #8, with F1 scores of 0.677 and 0.660) or bathymetry and hillshades (model #4, F1 score of 0.641). In order to test the contribution of the input layers to the model predictions, a series of additional model runs were carried out using both encoders but without implementing the semi-supervision -which the previous results reveal to be relatively erratic, and increasing the number of labels to 100 per class, to gauge the effect of boosting label number to model performance.

The results of this series of tests are presented in Table 3 , model numbers #19 to #40. As expected, the scores show an overall improvement caused by the increase in the number of labels from 50 to 100. While the time effort required to create labels for the classes is doubled (from 500 to 1000 labels in total), the improvements are significant, up to ~0.19 points, i.e., from 0.641 to 0.833 when comparing the F1 scores of the best VGG13 bathymetry and hillshade results (#4 vs #29).

Once again ResNet50 runs are slightly more successful in evaluation metrics compared to VGG13, with ResNet50 scoring higher in F1 8 times out of 11 model runs.

The exploration of the usefulness of each input layer in model performance provides strong indications that the hillshades are the most valuable set of layers for a correct prediction of morphological classes. Models that utilise hillshades have consistently higher scores than those that do not (cf. for example models #21 with #31 or #33 with #37, Table 3 ). The use of hillshades alone provides very good results (model #26 F1 score: 0.774), although the combination of layers with different azimuths is essential and a single hillshade is insufficient to produce an accurate map. Overall, the combination of the other derivatives alone or with bathymetry leads to substantially inferior predictions, with Precision scores consistently under 0.59 and Recall scores only slightly better. Aspect functions and VRM do not seem to provide useful information to the models, on the contrary their addition is detrimental to their performance. For example, bathymetry as input layer alone (#19 and #20) contributes to a better score than bathymetry combined with aspect functions and VRM (#33 and #34). While the bathymetric position indexes improve the predictions of the bathymetry baseline, they do not seem to enhance significantly the performance of the hillshade layers, with oscillating results when comparing the “HS full” baselines (models #25 and 26) and the “BPI + all HS” (models #31 and 32). The only layer that does improve the predictions of the hillshades alone is the bathymetry, with model runs #29 and 30 presenting the highest scores obtained in this study (VGG13, Precision 0.845, Recall 0.850). Visually comparing the map outputs of hillshades alone against bathymetry-supported hillshades shows improvement in score metrics obtained by the latter as reflected in the outlook of the map ( Figure 5 ), although the crispness of the boundaries is somewhat diminished, creating more “padded” class interfaces and generalisations.

The effect of the combination of all the input layers is given in runs #39 and 40, where the scores are only slightly superior to the model runs of the hillshades, but inferior to the bathymetry and hillshades runs. Overall, Large Ridge, Plane and Fissured (rock) are the three most successfully identified classes by all 10-class models, with an average F1 score of 0.884, 0.772 and 0.716 respectively. Corrugated (sediment), Hummocky (sediment) and Layered (rock) score instead the lowest across all models, with average F1 scores of 0.471, 0.466 and 0.423. The confusion matrices (see the Supplementary Material ) show that the prevalent misinterpretation is related to Type II errors (false negatives) where Layered (rock) is classified as Plane, Hummocky (sediment) as Depression and Corrugated (sediment) as Large Ridge. Probable causes for these misinterpretations are treated in the discussion.

Finally, models #13 to #18 ( Table 3 ) show the results of separate tests carried out to investigate the performance of the FCNNs with an increased number of labels. All the results show a substantial decrease in all the scores when moving from the 10 class to the 12 class problem, with Precision and Recall ranging between 0.442-0.567 and 0.451-0.584. Confusion matrices show a decline in accuracy in all classes, and not only those that were split. Depression (enclosed) and Bank (sediment) scored the lowest amongst the classes, showing that the separation from the original and more general Depression and Large Ridge classes (10 class division) weakens the training.

The use of several permutations and combinations of different input layers allows for an ensemble learning scenario to be leveraged. We have tested this hypothesis with a simple modal voting of FCNN pixel classifications for the 10 best performing models (both in terms of scores and visual quality), which produced an excellent map with an overall F1 score of 0.96 and class precisions and recalls superior to 0.87. The results and full map are presented in Table S1 and Figure S1 in the Supplementary Material .

In the breakdown of evaluation metrics for each class ( Table S1 in Supplementary Material ) the three most recurring weakest predictions are linked to the classes Corrugated (sediment), Hummocky (sediment) and Layered (rock). Coupling the observations of class type misinterpretation (see Results) and the qualitative assessment of the map outputs has led to the identification of three main types of errors or uncertainty.

Misclassifications linked to liminal spaces between classes is the first type of ambiguities we discuss ( Figures 9A, B ). This misclassification is reflected in the significant confusion between Layered (rock) and Plane or Hummocky (sediment) and Depression. Stratified, gently dipping bedrock possesses significant extents of planar features within them (bedding planes), that transition into fine elongated and often isolated ridges. This texture is sometimes misidentified as Plane, but in unlabelled data can also be observed as Layered (rock) in areas of sorted bedforms, that possess similar geometry. A similar case is provided by the Hummocky (sediment) class, which includes the occurrences of drumlins (oval shaped, glacial-flow aligned, moraine hills formed beneath fast-moving ablating ice flows). The drumlins are surrounded by depressed areas, the “connecting surface” between the high relief landforms. Models tended to confuse the proximal interconnecting surface as Depression instead of “drumlin”, leading to the lower score. In defence of the networks, it is often very difficult even for a geomorphologist to find the “correct” place to draw a boundary to define a landform (Smith and Mark, 2003). One major reason that labelling was carried out by a single expert, was to try to achieve maximum consistency in delineation, as another geomorphologist might introduce subjective bias and training conflicts for the networks. Moreover, complex terrains or where class assignment felt ambiguous were deliberately not labelled, leaving effectively the model to decide. We have stressed in the Methods section that good care was taken in the definition of distinctive semantic classes, however these errors indicate that morphological textures form part of a spectrum that is fundamentally difficult to compartmentalise (e.g., at what scale and configuration does a corrugation become a hummock or vice-versa)?, and the shortcomings of the FCNNs are at least partly by-products of natural variability and the inability of a set of classes to fully capture it. Without using a more complex set of classes and fuzzy classifiers it is not possible to treat any existing terrain variation.

The second type of ambiguity is related to scale ( Figures 7 , 8 , 10 ). Our classifications included the class Large Ridge (or Bank (sediment) and Relict ridge), which can be significantly bigger than other terrain or landform classes. This factor of scale ambiguity was introduced wittingly into the models, as we wanted to explore the “style” and ability of the networks to disentangle the problem of multiscale classification, which is very common in geomorphology and habitat mapping. If restricted to create a single map layer with a small number of classes, the human mind would prioritise the assignation of a class depending on what they think is the most important attribute to classify. So, for example, a large moraine which is covered by a boulder field might be preferentially mapped as “moraine”, even though both classes identify a correct characteristic of the ground. The hierarchical nature of BTM (Walbridge et al., 2018; Goes et al., 2019; de Oliveira et al., 2020) perpetuates this problem. In our results, this multiscale ambiguity is well reflected in the misclassification of Corrugated (sediment) as Large Ridge; corrugated surfaces such as smaller dunes or sorted bedforms occur extensively on the shelf and can overprint larger features, such as sediment banks or large moraines. The networks preferentially choosing the classification as Large Ridge might reduce the scores in the evaluation metrics but do not technically produce a wrong interpretation, rather a partial one. In some instances (e.g. the on shelf edge, see Figure 10 ), model predictions have dissected longer wavelength dunes (i.e. large underlying landforms) interpreting them partially as Large Ridges and partially as Corrugated (sediment), where the superficial sorted bedforms are more pronounced. Class prioritisation seems to be dependent on the way the model has learned the classes and boundaries, which in turn depends on adjusted weights and biases the model has learned during model training. However, understanding the individual activations and the internal workings of the neural network would require a study of class activation maps or the visualisation of deconvolutional layers (Noh et al., 2015).

Finally, a third type of recurring errors is connected to an inherent problem of the input layers: namely artefacts. MBES data can present many type of artefacts mostly caused by the limits of the instrument, the motions of the survey vessel (dynamic systematic errors), poor tidal or water sound velocity control causing vertical shifts and sound refraction. These artefacts are difficult to eliminate completely and a common obstacle in automated marine mapping (Lecours et al., 2017). Artefacts are recurrent in the extensive INFOMAR MBES bathymetry dataset, which is a combination of data from hundreds of different surveys with an array of vessels and survey operators, acquired with different (improving) instrumentation, in the space of about 25 years. The topographic variability introduced may consist in pixelation (salt-and-pepper effects), undulation along the swath, striping effects and cliff-like edges, and the vertical difference is often comparable with real features at seabed (e.g. megaripples or furrowing) ( Figures 9C, D ). Additionally, our hillshades are particularly susceptible to this kind of “topographic noise”, as they are vertically exaggerated to enhance the visibility of faint terrain patterns, which diminishes considerably their effectiveness. While a study of the effect of artefacts was outside the scope of this paper, it is reasonable to affirm that much stronger predictions can be achieved with a “cleaner” dataset.

Morphological maps provide the backbone for seabed habitat mapping studies, with classifications commonly obtained using semi-automated techniques as OBIA, BTM or other GIS tools (Harris et al., 2014; Goes et al., 2019; Linklater et al., 2019; Arosio et al., 2021) that segment the seabed in discrete parcels subsequently classified on the basis of pixel group statistics or geometrical characteristics. While grounded on mathematical rules and granting replicability, these techniques lack flexibility (e.g. how to treat morphological exceptions or near-isomorphisms) and require a good measure of engineering. Moreover, rules applied in one seabed region do not necessarily work elsewhere, so each dataset might need to be treated differently. On the contrary, FCNNs can provide the flexibility needed to capture any instance of discrete landforms or terrain textures without requiring ad hoc segmentation protocols (OBIA) or formulation of classification rulesets (BTM).

A semi-quantitative assessment of the effectiveness of the FCNN predictions for habitat mapping can be made comparing bedrock or sediment texture substrates to existing maps. We compared the predicted FCNN “bedrock” classes (Fissured, Layered and Hummocky (rock)), with the bedrock substrate layer produced by INFOMAR and available on the INFOMAR portal (INFOMAR, 2022). In Figure 11 we take the models with best scores in “bedrock” prediction (model #29 and the modal vote map) and overlap the INFOMAR layer. We limit the comparison area to a subsection of the entire dataset (indicated in Figure 11C ), as parts of the INFOMAR layer are mapped at very low resolution (e.g. the areas in Figure 11C pointed by the red arrows), introducing further deviations, and in other zones the Hummocky (rock) class includes also rough glacial till substrate. The best comparison is provided by the modal vote map, with a total bedrock area of 2721 km² (INFOMAR = 2336 km²) and an overlap of 77%. Model #29 has a slightly better overlap (~78%), but has also a larger area mapped as rock (3276 km²). Most of this excess bedrock is caused by misinterpretation of Layered (rock) ( Figure 11D ), which is over-represented in the model (F1 score 0.58). These numbers have to be taken with a pinch of salt, the mapping approaches are different (e.g. in the INFOMAR dataset the fissures in the bedrock outcrops are given another class), at a slightly different resolution and using different input layers (the INFOMAR map relies abundantly on backscatter data). Nonetheless, there is a broad agreement between the two, and the FCNNs consistently predict bedrock where it has been effectively mapped (see Figure 11 ). Moreover they give further information on the texture of the bedrock, which can be useful for habitat predictions (Novaczek et al., 2017). A similar comparison can be made with submarine dune fields. In Figure 12 we compare the general location of submarine dune ridges extracted using semi-automated techniques and checked manually (Arosio et al., 2023) with the class Ridge (sediment) in the best performing models (#12 and the modal vote map). Once more the results show an overall agreement, with Ridge (sediment) predictions corresponding with dune field areas ( Figures 12A, B ). In some places the FCNN is more efficient in identifying subtler ridges (e.g. Figure 12F ), however in parts the related classes Large Ridge and Corrugation (sediment) were preferentially selected (e.g. Figures 12E, F ). The models show higher levels of confusion in the presence of trochoidal dunes ( Figure 12D ) that are often misclassified as Fissured (rock) indicating that the labelling is not effective enough to train for this particular morphological distinction.

This exploratory study has shown that FCNNs have considerable potential for the creation of large scale seabed landforms and terrain textures map, and that even with relatively modest human input the results can be satisfactory. A clear semantic class definition and label delineation (including numerous boundary cases) will improve the accuracy of the classification, while a more rigorous consistency in mapping scale will most likely reduce ambiguity. Our results show that the optimisation of derivative selection helps the model outputs, and a combination of hillshaded layers contribute substantially to prediction improvements. Further insights on the contribution of each layer could be obtained using techniques based on feature importance, as saliency maps (e.g. Simonyan et al., 2014). The ensemble voting map, which constituted the best outcome of these experiments, clearly shows the utility of using learnt biases on different subsets of input data, and that assembling predictions from several ‘weak’ learners outperforms a single ‘expert’ network, which is the premise of ensemble learning (Ganaie et al., 2022). For further work, several FCNNs could be trained concurrently on different subsets of input data, and a loss could be calculated based on the confidence of individual networks (Goyal et al., 2020; Zhou et al., 2021). The latter is akin to several Decision trees in a Random Forest in classical machine learning (Cutler et al., 2012).

From a habitat mapper’s perspective, the use of FCNNs can be successfully applied to seabed maps for morphological characterisation, and very good results and flexibility can be achieved provided the model is well trained and furnished with clean data. Very large scale mapping endeavours, as that presented in Harris et al. (2014), could be easily replicated and improved upon using FCNNs. Moreover previously trained models could be applied on the new datasets that are being collected and gathered for Seabed 2030. If a sufficient volume of labelled classes is cooperatively assembled in a “dictionary” and made publicly available, it could be used by the community to predict morphological classes across different datasets, improving upon map objectivity and inter-comparison. The time invested in creating such a dictionary would be considerable but worthwhile, as the FCNN method will be eventually better, quicker and easily repeatable compared to semi-automated or manual digitisations. We shared our labelled dataset on GitHub (https://github.com/BrandonHobley/geomorph_deep) as a starting point. While discrete computer power it is necessary, the code is open source and requires a relatively basic level of coding expertise to be run, allowing for a widespread adoption.

FCNNs have also their significant drawbacks. Firstly, they are essentially a blackbox whose internal workings are not fully understood. Secondly, labelling and training at one determinate pixel resolution is most likely not transferable to a different one. So having mapped at 25m/pixel our dataset is probably ineffective to map at 2m/pixel, and more ad hoc labelling will be required. Finally, at this stage of sophistication, FCNNs fail to recognize complex geomorphological processes, especially in cases of isomorphism, so human intervention is still required. This limitation is also caused by the input types themselves, as bathymetry-derived raster data alone are often insufficient (for human geomorphologists too!) to unequivocally identify seabed landforms. Only when different types of datasets (seismic lines, ground-truthing etc.) can be included in the predictions, will machine learning be useful for more complex seabed geological interpretations.