Earth observation (EO) data, derived from remote sensing and unmanned aerial vehicle (UAV), have been recently demonstrated to be essential tools for the ecosystem monitoring and habitat mapping, combining high technological and methodological procedures for applied ecology. However, research based on EO data analyses often tend to focus on image processing techniques, neglecting the development of a detailed sampling design scheme needed for an exhaustive habitat detection. This paper shows the results of a novel approach for mapping coastal dune habitats at a fine scale, using a supervised machine learning model, through the combination of vegetation plot sampling scheme, synergic use of multi-sensor spectral imagery (UAV-VHR) and environmental predictors (e.g., LiDAR), object-based image analysis, and landscape metrics analysis. Proposed approach was tested in a protected area, established to preserve notable habitats along the Italian Tyrrhenian coast. A detailed sampling scheme was designed and carried out during spring and summer of 2019, combining simultaneously UAV flight acquisition and field vegetation survey data, collected at high precision positioning. The calibrated classification model achieved an overall accuracy of 78.6% (standard error 4.33), allowing us to accurately classify and map five coastal habitats, according to EUNIS (European Nature Information System) classification, which were further verified through a fully independent validation field survey. Results demonstrate that VHR imageries, combined with specific field survey schemes, can be exploited to train classification models used for the detection of plant communities (i.e., meso-habitat) and plant species at local scale. Our findings demonstrate that UAV-VHR data is a valid tool to produce high spatial resolution information in sand beach ecosystems, giving ecology research a new way for responsive, timely, and cost-effective ecosystem monitoring.