Exploring the effects of habitat management on grassland biodiversity: A case study from northern Serbia

The LDA results facilitated a separation of the three habitat management types. The first axis of the discriminant analysis explained over 95% of the total dispersion, while the combina- tion of the first two axes accounted for 100% of the variability (Table 6). All variables were more closely related to sites managed through mowing. While a small overlap was detected between grazed and mowed sites, it was much broader between mowed and unmanaged sites (Fig 3). It is also worth noting that the unmanaged area was relatively uniform compared to others, as much greater variability can be observed in the results related to both M and G sites. Among the predictor variables, total vegetation cover and diversity of plant species had the highest factor scores on the first axis, and according to this decision tree model was trained only on data related to plants. The decision tree model was trained in a classification setting where the habitat manage- ment measure labels were used as the algorithm output and the remaining dataset was used as the input. By conducting post-training analysis, we were able to extract and rank features by their relative importance, allowing us to determine which plant species are most relevant for different habitats. Although several model were trained and used during our experiments— including Random Forest and XGBoost [61] Gradient Boosted Tree models—Decision Tree Classifier model with maximum tree depth set to 30 proved most useful. While the two ensem- ble models provided a slight increase in accuracy, we opted for the Decision Tree Classifier as

it is fully interpretable and provided good validation accuracy (71%). When applying this model, the data was split into training and validation sets at a standard 80/20 ratio. For all model hyperparameters, a combination of manual tuning and grid search was used to find the most optimal values. Data was not normalised as all tested models performed well when applied to raw data. Table 4. Community similarity for plants, pollinators, and birds at study sites (G–grazed, M–mowed, and C– control). Plants Pollinators Birds G C G C G C C 0.26 0.29 0.46 M 0.36 0.28 0.24 0.24 0.62 0.64 https://doi.org/10.1371/journal.pone.0301391.t004 Table 5. Results yielded by GLM analysis of the effect of habitat characteristics on the pollinator and bird species diversity within the study area. Estimated parameter Standard error t value p value Diversity of pollinators Intercept 5.52 1.33 4.16 0.000328*** Distance to the nearest tree (m) -0.0038 0.0019 -2.01 0.055 Trees outside sites -0.043 0.0259 -1.69 0.101 Diversity of birds Intercept -0.72 0.66 -1.08 0.29 Trees outside sites (%) 0.02 0.009 2.28 0.03* Vegetation height -0.01 0.0076 -1.67 0.11 Plant diversity 2.10 0.50 4.15 0.00035*** Significance is expressed as * p < 0.05, ** p < 0.01, and *** p < 0.001. https://doi.org/10.1371/journal.pone.0301391.t005 PLOS ONE Effects of habitat management on biodiversity PLOS ONE | https://doi.org/10.1371/journal.pone.0301391 March 28, 2024 10 / 19