Multi-dimensional characterization of pelagic habitats and the potential to detect ecological niches in a highly dynamic ecosystem

Abstract

This thesis deals with multi-dimensional data sets sampled in the pelagic zone of the North Sea. Unlike terrestrial or benthic ecosystems, there are no readily observable differences in the pelagic zone. It is therefore difficult to define habitats or ecological niches and to predict the response of an ecosystem or its community to anthropogenic or environmental pressures. However, this is essential for sustainable management, especially in the face of global warming and increasing anthropogenic impacts on coastal ecosystems, such as offshore wind farms (OWFs). The results of this thesis highlight the potential of machine learning to improve our knowledge of the processes that shape plankton communities, but also its limitations. Fully automated methods are presented to (a) classify in-situ plankton images and (b) detect differences in the pelagic zone based on physical and biological measurements. The potential for detecting ecological niches and anthropogenic impacts on a highly dynamic ecosystem such as the North Sea is discussed. In chapter I (‘Automatic plankton image classification—can capsules and filters help cope with data set shift?’) the potential for automatic classification of in-situ plankton images using a Capsule Neural Network (CapsNet) was investigated. Data Set Shift (DSS) in this case describes the problem of shifting plankton communities in both spatial and temporal dimensions. The CapsNet was less affected by DSS than a standard convolutional neural network (CNN), but it also had a lower overall recall, especially for rare classes. The CNN classifications were more affected by DDS, but were still sufficient to reflect the spatial distributions observed in the field, at least in the case of the more abundant groups. For rare classes, an alternative method called ‘top-3 accuracy’ is proposed to limit human effort while increasing the recall of individual target species to >95%. In chapter II (‘Automatic segregation of pelagic habitats’) an Autoencoder (AE) was used to detect patterns in a data set consisting of biotic and abiotic variables. Each variable contributed a single value to a multi-dimensional micro-habitat that was projected by the AE onto a two-dimensional plane. The projections were clustered and grouped into macro-habitats consisting of similar micro-habitats. The method consistently identified three distinct pelagic macro-habitats, a ‘surface mixed layer’, a ‘bottom layer’ and a ‘productive layer’. Distinct plankton communities were observed in the different macro-habitats. Furthermore, anthropogenic influences induced by an OWF were successfully identified. The predictive power of the variables varied between data sets from different cruises, an indication of the complexity of interacting factors shaping pelagic habitats. In chapter III (‘Identification of plankton habitats in the North Sea’) I investigated the potential for predicting ecological niches from high-frequency multivariate datasets in the North Sea. The combination of an AE and a density-based clustering algorithm detected several complex habitat patterns, but niche segregation of plankton species at the sub-mesoscale was likely superimposed by local hydrography. Upwelling-downwelling dipoles (in the following simply dipoles) induced by offshore wind farms can develop similar characteristics as naturally occurring frontal systems and improve local productivity. Although of limited applicability at the sub-mesoscale, the model demonstrated the capability for rapid automated processing of multivariate datasets, a key requirement for future research given the increasing amount of data available to marine scientists and the complex dynamics of the pelagic zone.