Insights into Salt Marsh Plant Community Distributions Through Computer Vision and Structural Equation Modeling

Community structure and dynamics are influenced by numerous abiotic and biotic factors requiring large datasets to disentangle, which are often difficult to obtain over the spatiotemporal scales necessary for meaningful analysis. The approach outlined here illustrates one potential solution to this problem by leveraging computer vision methods to gain accurate, in-depth community data from ~ 10,000 photographs of salt marsh plants across an elevation gradient at Sapelo Island, GA, USA. A convolutional neural network (ResNext101) trained to detect the 6 dominant plant species achieved high accuracy for all species, allowing mapping of high-marsh plant communities over gradients in elevation and pore-water salinity. To statistically analyze the high-resolution mapping data, we constructed a structural equations model using the generated data as informed by prevailing ecological theory for salt marshes in the Southeastern United States. Model fit to data was strong, with R 2 values for five of six plant species > 0.7. The distribution of the rare understory perennial Limonium carolinianum , however, was not accurately predicted by the model. Modeled effects of abiotic factors elevation and soil salinity were commensurate with the literature. Biotic interactions also largely conformed to ecological understanding of Southeastern marshes, but a potentially novel positive interaction between Borrichia frutescens and Batis maritima was observed. Overall, this approach shows promise as a method of efficiently generating and statistically analyzing community data for sessile species at scales not previously possible. This study contributes to a growing body of work developing integrated computer vision and big data techniques for ecological field work.