This innovation area aims to develop deep learning models and applications for monitoring the marine environment.
Ecological studies, which involve e.g. classification and statistical counting of species in an ecosystem, are challenging and time-consuming tasks in marine science. Efficient and reliable data-driven methods for automatic analyses of complex marine observation data are needed to ensure sustainable fisheries and harvest. Deep learning has the potential to automate and streamline the steps required for such studies, but few applications in this domain have been developed.
Visual Intelligence has developed several innovations which enable efficient and reliable deep learning methods for automatic analysis of complex marine observation data. For instance, research within this innovation area has resulted in novel deep learning methods, such as for:
These methods have been developed in close collaboration with user partner Institute of Marine Research.
Many of the challenges of using deep learning in this area are related to training data, as annotated data can be both scarce and expensive to obtain. There is also a need for explainable and reliable models, as trust becomes important when the output of these systems are intended as input to models designed for abundance estimation. Deep learning methods which estimate uncertainty are also needed in the marine science. The innovations mentioned above were developed with these challenges in mind.
For example, the semi-supervised method for detecting and classifying fish species reduces the dependency on annotated training data, while making efficient use of available annotated data. Our work on explainable marine image analysis not only enhances the accuracy of marine species detection and classification but also provide clear insights into the decision-making processes of these models.
Newly developed methods for uncertainty quantification, such asthe one incorporated in a pre-trained network for sandeel segmentation in echosounder data, allow for pixel-based segmentation results with associated uncertainty.
When developing deep learning solutions to address challenges that our partners in marine science face, it is important to transfer knowledge and methodologies across innovation areas. Our proposed methodologies synergize well with other work within this innovation area, as well as our other three innovation areas.
Our work on explainable marine image analysis has been validated on multiple marine image datasets, such as multi-frequency echosounder data from underwater environments and aerial imagery of sea mammals captured by drones. The explainability framework can therefore be applied to different marine datasets.
The method for quantifying uncertainty incorporated for sandeel segmentation can also be used across a variety of network architectures. This allows for potential transferability of methodologies across the innovation areas.