Subpixel classification of high-resolution satellite images to identify dune plants

The increase in spectral and spatial resolution offered by satellite imagery opens new opportunities in ecosystem monitoring over broader scales. Combined with advances in machine learning algorithms, like the spectral unmixing methods, it is possible to extract information on features smaller than the pixel size, as previously demonstrated for the case of saltmarsh plants. The present study focuses on transferring subpixel classification algorithms to mid-latitude coastal dunes, a significantly more challenging environment in terms of plant size and density, as well as in terms of complexity and heterogeneity of the existing species. To this aim we obtained WorldView2 imagery over the Ria Formosa barrier islands (South Portugal) during May of 2023 and collected data on dune plant distribution over three barrier islands during the same period. A total of 800 m2 over the foredune (toe to lee) were photographed during fieldwork, identifying a total of 32 plant species. Plant density distribution was assessed at the level of the pansharpened satellite image pixel and the data were introduced to the Random Forest Soft Classification (RFSC) algorithm for training and validation. The sensing ability of the classifier was tested considering different parameters (number of trees, split criteria) and assessing the performance for increasing number of classes, along with the importance of the 8 spectral bands for each class. The results of the analysis provide insights on the strengths and limitations of the RSFC method for the especially challenging environment of mid-latitude coastal dune habitats and provide a step forward in coastal ecosystem remote sensing and monitoring of these environments.   Acknowledgements: The work was implemented in the framework of the DEVISE project (2022.06615.PTDC), funded by FCT (Fundação para a Ciência e a Tecnologia), Portugal. K. Kombiadou and S. Costas also recognise the financial support of FCT through contracts CEECINST/00146/2018/CP1493/CT0011 and 2021.04286.CEECIND, respectively, and the support of national funds through FCT by projects LA/P/0069/2020, granted to the Associate Laboratory ARNET, and UID/00350/2020, granted to CIMA.