|
"The biological carbon pump (BCP) is a natural process where phytoplankton in the surface ocean use carbon dioxide (CO2) from the atmosphere to first build their cells and then transfer this fixed carbon to depth in the form of sinking, aggregated organic particles (commonly referred to as marine snow). This process plays a pivotal role in regulating the Earth's climate by naturally sequestering carbon for long periods in the deep ocean. Broadly, the BCP sequesters carbon if these marine snow particles can sink to depth before they are consumed by microorganisms (called remineralization), which releases CO2 back into the water. Despite its importance, we are currently unable to predict whether changes in climate will increase or decrease the overall strength of the biological carbon pump, primarily due to an incomplete understanding of the numerous sub-processes and mechanisms that govern marine snow sedimentation and remineralization. An underlying cause is the heterogeneity of individual marine snow: each particle represents a microscale ecosystem where sinking rates and remineralization are governed by complex physical, chemical, and biological interactions. Without rapid methods for classifying this heterogeneity and relating it to the physicochemical properties of individual particles, our predictive power is limited by the measurement of bulk particle properties, which are plagued by high uncertainty and high variability. With the widespread availability of novel approaches and imaging techniques, we are now collecting detailed particle-based data at unprecedented scales and frequencies. This has led to an explosion of empirical data on particle optical properties, hydrodynamic behaviors, sinking velocities, and microbial-particle interactions. However, it has also created a bottleneck in downstream analysis that limits our ability to fully harness these vast and complex datasets.
The proposed project leverages multifaceted datasets collected by a team of transdisciplinary researchers (biologists, chemists, physicists, engineers, modelers) and establishes meaningful collaborations in computer science, data science and geosciences. It aims to merge the rapid expansion of aforementioned, high-throughput techniques to obtain particle-centric data with an urgent need to develop novel data processing approaches powered by artificial intelligence and machine learning (AI/ML). It builds upon a well-established and productive team of researchers centered around an NSF-funded multi-institutional project (see below) led by Kay Bidle (Professor, Department of Marine and Coastal Sciences; expert in microbial/biological oceanography and marine snow). Over the past five years, this collaborative has collected image data of individual marine snow particles from both laboratory culture-based systems and various oceanographic field campaigns, which will form the basis of the proposed project. Specific goals of the project are to:
1. Feature extraction and segmentation - Develop semi-supervised and unsupervised learning image segmentation methods to expand on the set of features that characterize marine snow including size, roundness, aspect ratios, fractal dimensions (i.e. porosity) and probability density functions of optical densities within individual particles and in particle populations.
2. Multi-modal learning - Develop AI models that can leverage the heterogeneous data modalities and fidelities for improved marine snow classification and modeling.
3. Multi-scale learning - Develop highly expressive ML models that go beyond traditional statistical parametric techniques and classical image segmentation to accurately model the complex physics governing the relationship between high-resolution particle-level data and context data from the same cruises such as physiological states of phytoplankton populations, their infection rates, and sinking fluxes of carbon from sediment trap collections, as well as with large-scale ecosystem processes under varied environmental and physiological conditions.
These goals are highly valuable to and needed by those working with image-based, marine snow data across various platforms. We envision that activities will focus on data derived from instrument platforms using holographic (LISST-Holo) and shadowgraph imaging (FoSI) of marine snow from recent expeditions in the Northwest Atlantic and California/Oregon coast. We also envision opportunities to calibrate imaging data with lab-generated marine snow particles of known compositional make ups. Data from both platforms currently have segmentation issues that need to be addressed in order to accurately identify particles and elucidate associated properties. We also anticipate opportunities (and needs) to explore platform-specific image processing pipelines as it relates to threshold-based, optical detection of polymers that catalyze the formation of marine snow and hold/glue it together, as well as 3D phase reconstruction using mostly published ML methods as a starting point."
|
Sign in
to view more information about this project.
