Deep learning for habitat mapping: exploring Italian coastal dunes

Coastal dunes host diverse and vulnerable habitats within limited areas, requiring effective monitoring tools [1]. While satellites provide quantitative data, the limited spatial and spectral detail of available imagery hinders broad-scale uses [2]. Convolutional neural networks (CNNs), integrating spatial and spectral features, emerge as a promising approach [3]. Thus, the aim of this study is to assess whether CNN-based methods can provide a solution for habitat mapping on coastal dunes at a broad national scale, using freely available RGB imagery from Google Earth and ground truth data from three EUNIS habitats (N14, N16, N1B). A pilot study was conducted in Collelungo (Tuscany, Italy). Ground truth data were collected in 2024 using 4 m2 squared plots, which were assigned to EUNIS habitats applying the EUNIS Expert System. An RGB image for the site was acquired from Google Earth with a spatial resolution of 0.2 m. To optimize the CNN, multiple U-Net configurations were tested, varying parameters such as the number of layers, filters, and kernel size. Different combinations of input variables were also assessed: the RGB image, a digital surface model, distance from the shoreline, and a NIR band from a regional orthophoto. Performances were compared in terms of loss, precision, recall and F-Score. To assess the transferability of the method, additional ground truth data were collected from sites distributed across eight Italian regions, representing broader habitat variability. For each site, one RGB image was selected from Google Earth with a date as close as possible to the field sampling period and the best-performing model from the pilot study was applied. Preliminary results suggest the potential of CNNs for supporting large-scale mapping of coastal dune habitats using freely available remote sensing data, contributing to the monitoring and conservation of these fragile ecosystems.