| Abstract [eng] |
In Lithuania, biologically valuable river valley grasslands are increasingly exposed to environmental pressures, as changes in agricultural practices and habitat fragmentation threaten the persistence of these ecosystems. Given the rapid decline in the extent of these habitats, their conservation, management, and monitoring require a better understanding of vegetation distribution patterns and the environmental factors shaping them. A spatial approach provides broader insight into habitat structure and its relation-ship with hydrological conditions. This study integrates hydrological modelling, field surveys, statistical analyses, and machine learning methods to investigate the spatial relationships of grassland vegetation in the valleys of the Nemunas, Nevėžis, and Šventoji Rivers. The objectives were: (1) to determine flood duration in the investigated river valley sections; (2) to assess the diversity and species composition of grassland communities; (3) to identify the influence of flooding and other key ecological factors on vegetation distribution patterns; and (4) to develop predictive maps of grassland vegetation distribution. For this purpose, 30 years of hydrological data and a high-resolution digital elevation model were used to generate a flood duration model, which provided the basis for field data collection. Grassland habitats were classified using expert systems, ecological differences among habitats were assessed statistically, and a random forest (RF) model was applied to predict habitat distribution and species composition. Nine flood duration zones (0–365 days) were distinguished during the study. Based on 66 study plots, 165 plant species and seven EUNIS grassland habitat were identified, including a habitat with intermediate floristic characteristics distinguished for the first time in this research. Grassland vegetation distribution patterns in river valleys were primarily determined by hydrological factors, particularly soil moisture and flood duration. The predictive accuracy of the random forest model reached 64.7% at the habitat level and 25% for species composition. The predictive maps revealed distinct spatial structures of grassland habitats across the study areas. This study is significant in several respects. Methodologically, it demonstrates that reasonably good predictive accuracy at the habitat level can be achieved even with a relatively small number of field plots. In addition, flood duration was confirmed as an appropriate quantitative indicator in vegetation studies. The predictive maps developed in this study may be directly applied to habitat inventory and monitoring, thus reducing the costs of field surveys. |
