Pixels and Patterns in the Sand: Decoding Coastal Dune Habitats With Multi‐Resolution CNNs

Le résumé fourni par la source

ABSTRACT Aims Mapping coastal dune habitats through remote sensing is crucial for biodiversity conservation, but it is challenging due to the small and fragmented nature of habitat patches relative to the spatial resolution of available imagery. Convolutional neural networks (CNNs), by leveraging both spectral and spatial information, offer a promising solution, but their application to coastal dunes remains limited. This study evaluates CNN performance for coastal dune habitat mapping using remote sensing data with varying spatial resolutions, testing how spatial resolution influences mapping accuracy and assessing the effect of including additional spectral bands beyond RGB. Location Coastal sand dunes in Maremma Regional Park, Tuscany, Italy. Methods Ground truth data were collected in 4‐m 2 plots and supplemented with photo‐interpreted patches, representing five classes: shifting dunes (EUNIS habitat N14), dune grasslands (N16), dune scrubs (N1B), bare sand, and sea. Four remote sensing datasets with varying spatial resolution were used: unmanned aerial vehicle (UAV; 0.02 m), airborne (0.20 m), Google Earth (0.30 m), and WorldView‐3 imagery (0.40 m). For each remote sensing dataset, one CNN was trained on RGB imagery and, where additional spectral bands were available, a second CNN was trained on multispectral imagery. Results Most maps achieved high accuracy, confirming the effectiveness of CNNs for habitat mapping. Accuracy of RGB‐derived maps declined with coarser resolution, from UAV (88%) to airborne (86%), Google Earth (80%), and WorldView‐3 (38%). Including additional spectral bands had varying effects on accuracy (+6 percentage points for UAV, +7 for airborne, +35 for WorldView‐3). Among habitats, dune scrubs were mapped most accurately, whereas shifting dunes were often confused with bare sand. Conclusions These findings underscore the value of very high‐resolution imagery for habitat mapping with CNNs, suggesting that even Google Earth imagery can support broader scale applications when UAV or airborne data are unavailable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.