Fully convolutional neural networks applied to large-scale marine morphology mapping

Abstract

In this study we applied for the first time Fully Convolutional Neural Networks (FCNNs) to a marine bathymetric dataset to derive morphological classes over the entire Irish continental shelf. FCNNs are a set of algorithms within Deep Learning that produce pixel-wise classifications in order to create semantically segmented maps. While they have been extensively utilised on imagery for ecological mapping, their application on elevation data is still limited, especially in the marine geomorphology realm. We employed a high-resolution bathymetric dataset to create a set of normalised derivatives commonly utilised in seabed morphology and habitat mapping that include three bathymetric position indexes (BPIs), the vector ruggedness measurement (VRM), the aspect functions and three types of hillshades. The class domains cover ten or twelve semantically distinct surface textures and submarine landforms present on the shelf, with our definitions aiming for simplicity, prevalence and distinctiveness. Sets of 50 or 100 labelled samples for each class were used to train several U-Net architectures with ResNet-50 and VGG-13 encoders. Our results show a maximum model precision of 0.84 and recall of 0.85, with some classes reaching as high as 0.99 in both. A simple majority (modal) voting combining the ten best models produced an excellent map with overall F1 score of 0.96 and class precisions and recalls superior to 0.87. For target classes exhibiting high recall (proportion of positives identified), models also show high precision (proportion of correct identifications) in predictions which confirms that the underlying class boundary has been learnt. Derivative choice plays an important part in the performance of the networks, with hillshades combined with bathymetry providing the best results and aspect functions and VRM leading to an overall deterioration of prediction accuracies. The results show that FCNNs can be successfully applied to the seabed for a morphological exploration of the dataset and as a baseline for more in-depth habitat mapping studies. For example, prediction of semantically distinct classes as “submarine dune” and “bedrock outcrop” can be precise and reliable. Nonetheless, at present state FCNNs are not suitable for tasks that require more refined geomorphological classifications, as for the recognition of detailed morphogenetic processes.

Seabed classification of multibeam echosounder data into bedrock/non-bedrock using deep learning

“…It can be thus inferred that U-Net can effectively generate all the necessary data representations from the bathymetry data alone. These findings differ from those presented in the study conducted by (Arosio et al, 2023) where a combination of bathymetry and hillshade data sources yielded DL models with the best performance. This difference in results may stem from the specific classification tasks of each study, in fact while Arosio et al (2023) aimed to identify various seabed morphological classes, including distinct rock textures, we focused solely on bedrock/nonbedrock separation.…”

“…These findings differ from those presented in the study conducted by (Arosio et al, 2023) where a combination of bathymetry and hillshade data sources yielded DL models with the best performance. This difference in results may stem from the specific classification tasks of each study, in fact while Arosio et al (2023) aimed to identify various seabed morphological classes, including distinct rock textures, we focused solely on bedrock/nonbedrock separation. Furthermore, our study utilized a highresolution, expert-generated map for annotation, in contrast to the limited annotated data employed by Arosio et al (2023).…”

“…This difference in results may stem from the specific classification tasks of each study, in fact while Arosio et al (2023) aimed to identify various seabed morphological classes, including distinct rock textures, we focused solely on bedrock/nonbedrock separation. Furthermore, our study utilized a highresolution, expert-generated map for annotation, in contrast to the limited annotated data employed by Arosio et al (2023). Although the disparities in our classification objectives and available data may account for our differing results, further research is essential to fully unravel the underlying causes of these discrepancies.…”

“…In the marine environment it has been successfully applied to underwater images for the study of the behaviour of marine sponges (Harrison et al, 2021), for the segmentation of fish species (Nezla et al, 2021), underwater mineral images (Wang et al, 2022) and underwater litter (Wei et al, 2022). Notably, recent research has demonstrated the potential of U-Net for deriving seabed morphological classes, including a hard-substrate class comprising bedrock outcrops, using high-resolution bathymetric data alone and a limited amount of labelled data (Arosio et al, 2023). This network has also been tested for onshore bedrock mapping using a digital elevation model (DEM) and cloud-based Landsat 8 data (Ganerød et al, 2023).…”