Coastal dunes are important habitats that provide a variety of ecosystem services (ecological, economic, coastal protection, etc.) and, as such, their monitoring is a priority for environmental protection (i.e., EU Directives). Eco-geomorphologic feedbacks between dune plants and coastal topography are fundamental to the self-organisation capacity of coastal dunes and a shift in community structure and composition (i.e., expansion of invasive species) can cause a domino effect, potentially crippling previously established system adaption mechanisms. It follows that monitoring dune vegetation is crucial, especially in protected areas and in fragmented and stressed dune environments. Even though recent improvements in spectral and spatial resolution of satellite imagery open new and exciting prospects for large-scale environmental monitoring, this potential is largely unused in dune ecogeomorphology, due to the challenges related with the small size and density of dune plants and the complexity and heterogeneity of the existing species. Machine learning techniques and subpixel classification methodologies, like the Random Forest Soft Classification (RFSC), have shown promising results in similarly challenging environments in terms of plant size and heterogeneity, with high accuracies in subpixel fractional abundance of marsh-vegetation species. Even though subpixel classification could improve monitoring biodiversity from satellite imagery, similar approaches have never been tested for dune environments. These challenges and gaps inspired the present work, built around the idea of testing subpixel classification methods for dune plant species identification using high-resolution satellite imagery. Here we present preliminary results from the application of RFSC to the western barrier of the Ria Formosa system (S. Portugal) using WorldView2 (2017) imagery and training/validation samples from UAV, along with the next steps planned to test the hypothesis that RFSC methods can be successfully used to identify dune plant species and to assess their predictive capacity and identify potential limitations.
