Multitemporal NDVI classification of coastal dune habitats of central Italy with Sentinel 2 data

Given the complex mosaic of plant communities coexisting in a relatively small space, coastal dune ecosystems have long represented a challenge for habitat mapping and multi- temporal monitoring applications. However, being one of the most threatened ecosystems on Earth, cost-effective ways to conduct recurrent observations over vast areas are urgently needed. In this context, the increasing high spatial and spectral resolution of open access, remotely sensed data is promising for monitoring. We tested the habitat mapping strength of a multitemporal Sentinel 2 imagery dataset for a sector of the Tyrrhenian coast (Lazio) at two detail levels: i) main land cover types (woody vegetation, herbaceous vegetation and sparsely vegetated areas) and ii) complex of habitat types (EU Habitats sensu 92/43/EEC ). Spectral analysis was based on the Normalized Difference Vegetation Index (NDVI) temporal variation, as proxy of vegetation phenological proprieties, and classification on random forests method. All the processing chain was carried out with the open source products of the European Space Agency (sen2cor processor, SNAP toolboxes). Then we assessed through aerial images or field floristic surveys. We identified 3 land cover types with 80% general accuracy and over 60% producer accuracy, and 3 habitat complexes classes with 73% general accuracy and a weak (53%) to high (80%) producer accuracy, depending on the habitats complex. In this study, multitemporal Sentinel classification resulted an effective instrument for mapping an highly fragmented coastal dune system. This is an encouraging perspective for extending the open source earth observation monitoring techniques (from wide areas) to the local scale of the Italian and Mediterranean sand dunes landscapes in highly fragmented territories.