Habitat classification of central European sandy grassland using machine learning and field survey data

Habitat degradation and destruction have been identified as the main causes of the global biodiversity crisis. Habitat restoration-the process of aiding the recovery of habitats and bringing them to a former, more natural state-has been a topic of extensive policy initiatives and practical efforts. However, many restoration projects still fall short of achieving significant ecological improvements, often due to limited resources and a gap between scientific knowledge and on-the-ground implementation. We tested a machine learning method-the gradient boosting model-in a local study on a protected grassland in northern Serbia, currently under restoration, to assess its potential to correctly classify habitat types, based on ground-truth data. Model performance was evaluated across multiple configurations and feature selection steps. Results showed that our approach achieved high classification accuracy, using a reduced subset of species, indicating that a small number of taxa carry most of the predictive power. The identified key species, such as Festuca rupicola (Furrowed fescue) and Stipa pennata (European feather grass), aligned well with known diagnostic indicators of Central European semidry and steppe grasslands. Overall, our findings demonstrate the usefulness of machine learning-supported classification as a complementary tool for habitat monitoring, while highlighting the need for methodology testing in different settings.