Areas occupied by grasslands, arable land and natural succession are continuously changing over time. Due to these dynamics, reliable and comprehensive data on their distribution in Croatia remain only partially available. Remote sensing may offer an effective solution for large-scale identification and monitoring of non-forest plant cover, including the spread of grassland succession on abandoned grasslands. The aim of this study is to assess the potential of combining Sentinel-2 imagery with Discriminant Analysis (DA) and Classification Trees (CT) to differentiate categories of non-forest plant cover. Research was conducted on the south-eastern slopes of Medvednica Mt., within the wider Goranec Protected landscape. 59 polygons of interest were classified as grasslands, grassland successions, and annual crops. Centroids of a 10 m grid were used for sampling four spectral bands (blue, green, red, NIR) acquired between March and October 2024. Descriptive statistics highlighted distinct seasonal trends in May, July, and October in the spectral bands, which were selected as predictor variables. Overall, the CT model had higher accuracy (82%) than DA model (70%), but when compared at the category level, DA had higher accuracy for succession (86%), while CT for grasslands (90%) and arable land (48%). Consensus on accurately classified cases between the models was lowest for arable land (31%) and highest for grasslands (70%), with DA showing a tendency to overestimate succession, while CT overestimates grasslands. Initial results confirm the potential of Sentinel-2 data for classification of non-forest plant cover providing a cost-efficient tool for land monitoring and conservation planning, although additional testing and model tuning are needed.
Levačić et al. (Wed,) studied this question.