The grass is always greener: remote sensing and machine learning reveal severe and enigmatic loss of a salt marsh foundation species

Background/Question/Methods

: Salt marshes are globally important ecosystems which house critical services such as habitat for many animals, water filtration for upstream inputs, and carbon sequestration. Due to the impacts of global change, many of these ecosystems have been lost or are trending toward collapse due to shifting communities and sea-level rise. There have been previous attempts to broadly quantify salt marsh communities, specifically the grasses which serve as foundation species such as Spartina alterniflora and Spartina patens, the latter of which is less tolerant of sea-level rise. These grasses are ubiquitous and distributed throughout many salt marshes on the East Coast of the United States, yet they perform different functions in the ecosystems they cohabitate, indicating their shared importance. Further, there is a lack of research using high-resolution geospatial imagery to quantify fine-scale changes across the broad spatial extent of many ecosystems. Our research seeks to evaluate shifts in target salt marsh communities using available imagery over 13 years and to highlight the importance of funding for geospatial data programs. Specifically, we developed a novel, easy-to-use method of classifying geospatial data with supervised machine learning that enables a user to iteratively build datasets based on their own ecological knowledge.

Results/Conclusions

: Our methods allowed us to characterize immense regions ( >7000ha for each of two geospatially proximal areas) of critical salt marshes on the New Jersey coast and to evaluate fine-scale (1m × 1m) community transformations in response to global change for several years of historical imagery. We achieved strikingly high classification accuracy for all years, upwards of 98.75% correct for our final model in 2019. We were also able to achieve the highest classification accuracy in our models for Spartina patens versus other land classes used, at 100% for 2019 and >95% for all other years. These results confirm the viability of our classification methods for broad-scale community characterization. Analyzing classifications from 2006 to 2019, we found vastly different patterns of community response in our target marshes. One marsh appeared to experience very little change while the other experienced an incredible 81.17% reduction of Spartina patens in favor of Spartina alterniflora, potentially devastating the range of services able to be provided by this ecosystem. These results signify the importance for future broad-scale ecological studies to be inclusive in nature, so that the full range of global change response for a given ecosystem-type can be captured.