Oct 26, 2018

Identifying Tiny Organisms with Giant Impact on Climate

Scientists are experimenting with machine learning to automate identification and classification of plankton — tiny marine organisms that provide much of the air we breathe. An essential part of life, plankton form the base of the aquatic ecosystem, produce more than half of the oxygen on Earth, and remove carbon dioxide from the atmosphere. Plankton provide clues about climate change, harmful algal blooms, and more.

“We’re optimistic that machine learning technologies will vastly increase our ability to understand what’s out there and how it’s changing,” said Jules Jaffe, a research oceanographer at the Scripps Institution of Oceanography at the University of California San Diego.

One Billion Images Captured

Jaffe’s Scripps Plankton Camera System consists of two underwater microscopes that can detect objects as small as tens of microns and capture high-resolution images of hundreds of thousands of individual plankton each day. The dark-field cameras are deployed “in-situ,” able to observe plankton undisturbed in their natural setting, off the Scripps Pier. Taking pictures eight times per second 24 hours per day for the past three years, the cameras have captured more than one billion images. Pictures are published online, where scientists, students, and the public can explore and tag the data.

But before the images can be used for scientific analyses, the plankton must be identified and classified. “With volume like that, a human can't reasonably look through all of those images,” said Eric Orenstein, a post-doctoral researcher working with Jaffe and professor Peter Franks. “To get some sort of environmental signal out of that, we would like to start implementing machine learning in real time.”

Jaffe and his team sample more than 100 times the plankton data that can be obtained by traditional methods of human collection, which usually involves scientists scooping up plankton in nets and viewing individual samples under a microscope.

Machine Learning Identifies Plankton

To implement machine learning, Jaffe’s team first establishes training sets to teach a computer algorithm what to look for. For example, Kasia Kenitz, another post-doctoral researcher, is training the computer system to identify diatom colonies. A kind of plankton, diatoms are easy to recognize because of their unique cell walls, which form intricate, striking patterns — microscopic works of art. Kenitz feeds a support vector machine with plankton images and 79 characteristics that determine whether each image is a diatom colony or not, training it how to perform the classification automatically. “So far, the accuracy of the classifier is about 90 percent,” she said.

To conduct the research, Jaffe’s team implements many-layer convolutional neural networks -- deep learning networks that identify and analyze visual imagery and are inspired by the human brain -- such as AlexNet, ResNet, and SqueezeNet. They use high-performance computers called Flash I/O Network Appliances (FIONAs) that are optimized to communicate data over fiber optical networks at 10 to 100 gigabits per second (Gbps), enabling big-data applications like Jaffe’s with graphics processing units (GPUs) to analyze data more quickly. Jaffe’s team uses UC San Diego’s Science DMZ (Prism) to move data on campus; and CENIC’s High-Performance Research network for inter-campus data sharing. FIONAs were developed by the National Science Foundation-funded Pacific Research Platform with principal investigator Larry Smarr.

Finding Climate Solutions Faster

Speeding up scientists’ ability to understand plankton could ultimately help prevent harmful algal blooms and tackle climate change. Toxic algal blooms, known as “red tides,” can impact human health, marine life, and nearby economies. In August, a bloom in Florida prompted the governor to declare a state of emergency after piles of dead fish washed up on beaches and people experienced respiratory illnesses.

“There has been a global epidemic of toxic algae,” Jaffe said. “In California, we have our own version of these ... so there are regular samplings to make sure paralytic plankton are not in the waters.” Climate change and warmer water temperatures could lead to more frequent algal blooms. Even small changes in the growth of plankton may affect atmospheric carbon dioxide concentrations, which could speed up the warming of global surface temperatures. Carbon dioxide emissions — like the kind that cars produce — are absorbed by plankton.

Advances made by Jaffe and his team will benefit both the fields of computer science and biological science, enabling improved classification algorithms for machine learning and vast new data streams for plankton ecology. With success, underwater microscopes deployed around the world could be connected by computers that are analyzing and comparing plankton changes throughout the oceans in real time.

The ability to rapidly and automatically find patterns in plankton ecology could drastically speed up observations of climate change challenges and solutions. In recent predictions of technologies that could change the world, IBM claimed underwater microscopes powered by artificial intelligence could transform researchers ability to monitor ocean health. “Plankton are resources that directly impact human health,” Jaffe said. “We want to replace the technologist sitting at the microscope with a machine learning algorithm.”

Related blog posts

From the Ground to the Stars: Critical Big-Data Research in Africa

Learn about critical data-dependent research collaborations currently taking place in the nations of Africa, all of which require integration into the global high-performance network and cyberinfrastructure environments.

Get Your Cybersecurity Program Up and Running with the Trusted CI Framework

Learn about the Trusted CI Framework, developed by the NSF Cybersecurity Center of Excellence and created to help your scientific research facility overcome the obstacles to standing up a cybersecurity program for its cyberinfrastructure.