Natural Resources Science and Management (NRSM) PhD candidate Daniel Xie is exploring new approaches to modeling carbon in forests and aquatic ecosystems. Along with his advisor, Assistant Professor Chad Babcock, he is reimagining current terrestrial carbon modeling methods while training an AI system to accurately identify aquatic carbon storage. Xie took a moment to describe his research questions, their role in spatial modeling, and future opportunities for carbon accounting methods.
What research question are you asking?
Broadly speaking, my research question is, “How can we improve methods for modeling green and blue carbon?” As that suggests, I have two primary research directions: one focused on forest carbon (green carbon) and the other focused on seagrass meadows (blue carbon).
More specifically, my green carbon research focuses on answering, “What are the best methods to reliably predict carbon over different spatial scales?” My blue carbon research focuses on answering, “Can AI identify seagrass patches across coastal ecosystems?”
Why is this important?
As climate change impacts intensify year by year, the value of carbon storage by natural ecosystems becomes increasingly clear. As such, it is important to be able to assess carbon in terrestrial and aquatic ecosystems.
Forest carbon modeling is a well-studied topic. A common issue across forest carbon models is a loss of accuracy when predicting at changing spatial extents. For example, a state-scale model trained on statewide data may not be great at predicting the carbon content of a small local forest. I’m aiming for a more flexible model framework that remains useful for studying smaller forests of interest.
Despite their small global area, seagrass meadows capture 10% to 15% of all carbon in the oceans. Due to their aquatic nature, studying seagrass meadows often requires diving to conduct field sampling. Taking photos from the bottom of a boat (or aquatic drone) is an alternative to such difficult and expensive methods. However, the sheer amount of photos generated is difficult to transform into usable data. Unless this is resolved, the photo method is no less inaccessible than diving.
How are you approaching the topic?
For my green carbon research, I am developing a scalable carbon model that remains reliable across varying spatial extents. To be exact, I am using a two-stage Bayesian Hierarchical Spatial Model. With this approach, the first-stage model makes predictions of a large spatial extent (in my case, the state of Minnesota), then the second-stage model makes more refined predictions for any arbitrary smaller area of interest (AOI). This second stage requires additional training data from the smaller AOI, so I will also test the model’s sensitivity to differing availability of stage 2 training plots.
For my blue carbon research, I am capturing photo data of seagrass meadows by boat and training AI classification models to identify seagrass within photo data. From about 20,000 total photos, I identified about 4,000 containing seagrass. Then I annotated the seagrass photos. I use those annotated photos to train an AI seagrass classification model and test how well AI can identify seagrass.