AI Oceanography – The ‘Fourth Industrial Revolution’ Comes to Marine Science

Artificial Intelligence – in 2019 the phrase is practically inescapable. From self-driving cars to AI-enabled smartphones; expert game-playing computers to AI weapons, there has been an explosion in the performance and applicability of machine learning techniques for a seemingly endless variety of tasks. Led by the likes of Amazon, Google, Apple and Samsung, companies around the world are revolutionising their businesses using a host of techniques driven primarily by recent developments in deep learning (large neural network) algorithms and steadily decreasing costs for computational power. Far from trailing behind the corporations, scientists across many disciplines are at the forefront of discovering new ways to exploit the powerful pattern-recognition and statistical reasoning enabled by modern AI. In the life sciences, AI is delivering insights into the human genome, predicting cancer development and allowing drug discovery to accelerate to a pace previously unimaginable.

Anthony1U-net AI for image classification and segmentation – detecting important features road awareness. Many developments in neural networks were made possible by research for self-driving cars. (https://neurohive.io/en/popular-networks/u-net/)

In the earth sciences, AI is improving climatic and ecological time series by imputing missing data and consolidating conflicting observations, correcting for bias and building better predictive models than previously possible. The longstanding and seemingly intractable geophysical research question of where and when the next major earthquake will next occur, may, for the first time, be soluble using neural networks. The promise of machine learning solutions to issues which would be impossible or extremely challenging using traditional methods is alluring, to say the least.

Marine science has its own unique challenges and uncertainties – whether in collecting data over vast spatiotemporal scales, tracking and isolating the effects of different water masses in highly dynamic systems or simply accessing the remote and often dangerous regions of interest. High-resolution, long-term in situ observational datasets are becoming possible thanks to pioneering projects like Argo, which now has around 4000 widely-dispersed autonomous platforms taking regular CTD & current velocity measurements in the upper 2000m of ocean. Biogeochemical Argo floats are also increasingly coming online, additionally enabling measurements of chlorophyll fluorescence, acoustic backscatter, dissolved oxygen and other key ocean characteristics. For such large datasets as the global Argo observational output, deep learning techniques can be used for resolving turbulent processes, sub-surface currents, air-sea fluxes and energy transport; interpolating measurements across areas with sparse coverage; improving model parameterisation; automatic quality control and more. ­­­­­­Even data-collection itself can be greatly aided by new autonomous ocean vehicles, with current AI research enabling intelligent autopiloting and adaptive sampling using ocean gliders and other AUVs.

Anthony2Currently active Argo floats (http://www.argo.ucsd.edu/About_Argo.html)

The primary driver of these impressive recent developments has been significant advances in deep learning architecture designs. Of these, convolutional neural networks (CNNs), which are ideal for detecting features in multidimensional data such as images, are most prominent. CNNs are networks of spatially-interconnected neurons – simple algorithms taking several inputs and having one output, a nonlinear function of the inputs and a bias value. By building up many layers of neurons and optimising the neurons’ output functions, behaviour emerges that mimics the structure of the human visual system. For arduous tasks such as the processing of acoustic spectrograms and subsea imagery, which is traditionally carried out by researchers with the help of undergraduate research assistants, CNNs can offer extraordinary improvements in both speed and accuracy. Where it is time-consuming to manually pick out sparse features in mostly uninteresting seafloor images, CNNs can intelligently segment images and output only features of interest for later identification. Moreover, if enough labelled data can be made available for the optimisation of a CNN, in addition to recognising the presence of an animal to be taxonomically identified, the network can be set up to distinguish between taxa and perform the identification itself.

Anthony3Deep learning detects fish in real-time from ROV imagery (https://www.saf21.eu/2018/04/26/workshop-on-uses-of-machine-learning-in-fisheries/)

A particularly exciting application of deep learning to biological oceanography, and one I am personally pursuing for my PhD project, is the use of CNNs to process microscope images. Phytoplankton are critical indicators of water quality, climate change and primary production, and can be devastating to marine and human life in Harmful Algal Bloom events. Despite their global significance, it is very difficult to sample and classify plankton at the scales needed for accurate representation of their diversity and distributions. My project will go some way toward addressing this issue by developing an open-source, low-cost instrument for high-speed, high-performance digital microscopy of microplankton. By enabling the wider oceanographic community to collect millions of images of phytoplankton in mixed, untreated seawater samples and make public their data, I aim to allow the development of robust CNNs which can automatically detect and taxonomically identify species of interest. My project forms part of the third cohort of the NEXUSS (NEXt generation Unmanned and autonomous Systems Science) Centre for Doctoral Training at the National Oceanography Centre Southampton, which is one of several programs designed to promote intense collaboration between computer scientists, engineers and oceanographers. Other exciting NEXUSS projects include the use of CNNs for seafloor image classification, the prediction of seabird migration patterns, the design of new ocean sensors and platforms, and the development of autonomous navigation and control for AUVs (including Boaty McBoatface!).

There is a significant global drive towards collecting and making available high-quality image databases for deep learning optimisation. The excellent Seafloor Explorer project by the Woods Hole Oceanographic Institution leveraged citizen scientists and distributed effort to generate an enormous training dataset for automated seafloor analysis. Other marine projects include the creation of a large online database of citizen scientist-labelled zooplankton images and the publishing of over 4 million pre-labelled phytoplankton microscopy images. As more data becomes available for its optimisation, AI will increasingly be able to improve oceanographic understanding, increase the scope of research and automate expensive and laborious manual tasks. In the near future, a global ocean observing system that measures and uses AI to interpret physical, geochemical and biological marine characteristics will also be able to move sensing platforms for optimal coverage/resolution. This new GOOS will rapidly communicate its findings to oceanographers over the internet, spot patterns in the enormous datasets it produces, and integrate with models to make predictions of future trends. This new era of oceanography will be possible based on the individual components currently under development.

Anthony4Most Seafloor Imagery is not this exciting! Citizen science projects building labelled datasets will enable AI systems to scan millions of images and detect living organims, as well as sedimentary/geological features, with higher precision and lower resource use than ever before. (https://www.seafloorexplorer.org/)

This post was written by Anthony Lindley, PhD Student, National Oceanography Centre/University of Southampton.

a.lindley@southampton.ac.uk

@AJWL27

 


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s