Petra Šímová, Jiří Prošek, Petr Klápště, Duccio Rocchini, Vítězslav Moudrý

Mapping and monitoring of Natura 2000 habitats (Habitat Directive 92/43/EEC) is one of the key activities to ensure the protection of natural habitats in Europe. Remote sensing can help to acquire high-quality maps of the distribution and conservation status of Natura 2000 habitats, for example through classifying multispectral data. However, due to the high number of habitats (classes) distinguished in the Habitat Directive, achievable classification accuracies for individual habitats in the context of a given landscape remain unknown. Moreover, although many recent studies have brought encouraging results in the classification of very-high-resolution satellite data such as Sentinel-2 or Rapid-Eye, spatial resolution in the order of several decimetres achievable with UAVs can be needed for distinguishing individual habitat patches in fine-grain landscape mosaics. In this study, we investigated the potential of UAVs for distinguishing eleven Natura 2000 forests, wetlands and grasslands. The study area (~ 20 km2) is situated in the heart of the Czech Republic, Central Europe. We aimed to assess the producer, user and overall accuracy of Random Forest classification, considering the importance of different phenological seasons (spring, summer), spectral resolutions of the camera (multispectral and RGB), predictor types (spectral, textural and object) and classification scheme (detailed habitats vs their aggregations). The highest achieved overall classification accuracy (Cohen’s Kappa) ranged from 0.71 to 0.77 and resulted from classifying multispectral data from both seasons. We obtained similar results from the spring season (0.67-0.76), whereas the isolated data from summer provided poor distinguishing capacities. Relatively good accuracies (0.65 to 0.75) were achievable even using a simple RGB camera when combining both seasons. In general, the classification of non-forest habitats was better than that of forest habitats. Spectral predictors (mean band values) played a crucial role in the classification, but including the object properties and texture (spectral variability) also improved the distinguishing capabilities.