Deriving benthic traits through deep learning methods

Through global warming, ocean deoxygenation is considered as a major concern since it consequently reduces the quality and the quantity of suitable habitats for marine life. Eutrophication plays a major role in its depletion which enhances respiration at different depths. Many species such as fishes, benthic worms or even plankton are affected by this phenomenon.

This study aims to get a better understanding of benthic worm species on the continental shelf of the Black Sea which is well known for high frequency oxic stresses. Our main objective is to map species through their biological traits (i.e. body length, burial depth, reproductive frequency…)  in order to assess their vulnerability towards environmental variations that occur at this location. 

Unfortunately, in the oceanographic field, one of the major issues is the sparsity of in-situ observations, especially when it comes to benthic biology. Therefore, we have decided to use a multivariate approach allowing us to use related datasets with significantly better spatial and temporal coverage. This multivariate approach is implemented using deep learning in order to get complete maps of traits on our domain. An adapted convolutional neural network allowing to capture non-linearities is used to reconstruct the traits repartitions. 

Thus, as an input for the neural network, we consider our traits dataset and environmental variables which are likely to enhance their reconstruction; Surface currents, particulate organic carbon, oxygen concentration and bathymetry are considered. A chosen period from 2008 to 2017 is selected. Traits datasets are located by stations (238) and were constructed through fuzzy coding and rescaled by their biomass. 

The neural network architecture is composed of an encoder and a decoder where the encoder considers a gappy and non-gridded dataset. The encoder uses a series of convolutional layers followed by max pooling layers which reduce the size of the dataset. The decoder does essentially the reverse operation by considering convolutional and interpolation layers. 

In order to avoid overfitting, the model has skip connections which ensure to keep information from the input dataset. For additional information please refer to Barth et al 2022. The model gives the reconstructed trait repartition and the standard error of the reconstruction.

This study will be helpful in the understanding of benthic traits repartition and will aim to link their patterns to environmental factors. This will help to get a deeper understanding of the ecological role and functions of this poorly known ecosystem. This work is carried in the frame of NECCTON European project.