A computer vision-based approach for estimating carbon fluxes from sinking particles in the ocean

Abstract

The gravitational settling of organic particles in the ocean drives long-term sequestration of carbon from surface waters to the deep ocean. Quantifying the magnitude of carbon sequestration flux at high spatiotemporal resolution is critical for monitoring the ocean's ability to sequester carbon as ecological conditions change. Here, we propose a computer vision-based method for classifying images of sinking marine particles and using allometric relationships to estimate the amount of carbon that the particles transport to the deep ocean. We show that our method reduces the amount of time required by a human image annotator by at least 90% while producing ecologically informed estimates of carbon flux that are comparable to estimates based on purely manual review and chemical bulk carbon measurements. This method utilizes a human-in-the-loop domain adaptation approach to leverage images collected from previous sampling campaigns in classifying images from novel campaigns in the future. If used in conjunction with autonomous imaging platforms deployed throughout the world's oceans, this method has the potential to provide estimates of carbon sequestration fluxes at high spatiotemporal resolution while facilitating an understanding of the ecological pathways that are most important in driving these fluxes.