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

preferences, and ecological connections, decision trees facilitate ecosystem management, bridging the gap between research and practical conservation efforts, ultimately supporting biodiversity preservation and ecological sustainability. In the field of computer science, deci- sion trees are artificial intelligence models which belong to the machine learning subbranch of the artificial intelligence field. They are not hand-made, but rather trained on our data in order to give us meaningful insights and predictions. From the inputs and outputs of setup decision tree model, thresholds for decision making inside the model are automatically learned to minimize error which in our case is classification loss. For this analysis, as described later, we used Python programming language. As the dataset used in this study exhibits class imbalance, several methods were explored to overcome this issue. The imbalance comes from the nature of the study, with one control set of observations compared to three for the other output classes (mowing, grazing). In the process of training our Decision Tree model, we tested various approaches to overcome the class imbalance of the data. Initially, we attempted using the Decision Tree classifier’s inherent feature to apply class weights, assigning more weight to the minority class. This was achieved by controlling the ’class_weight’ parameter in the ’scikit-learn’ library implementation of the Decision Tree Clas- sifier. In the model setup, we assigned three times the weight to the Control output class com-

pared to the other two classes. Unfortunately, this adjustment did not lead to improved performance. To overcome this problem, we explored both undersampling and oversampling techniques. Undersampling posed a risk of losing crucial information necessary for effective model learning due to the dataset’s limited size. To address the class imbalance, we employed the SMOTE [56] oversampling technique from the ’imblearn’ Python library. Given the restricted variety in the control data, obtaining more diverse and substantial data in the future would be advantageous in alleviating the class imbalance issue. We adopted the decision tree implementation from the “scikit-learn” [57] package and relied on the “pandas” [58] and “numpy” [59] libraries for data preprocessing. In addition, for exploratory data analysis in preparation for model training, “ydata_profiling” package [60] was adopted. Due to the tabular nature of our dataset, we opted for decision tree-based algorithms as they provide good bal- ance between performance and interpretability. These models were used as classifiers, whereby the training phase allows the utility of certain species for segregation of different habitats to be ascertained. During the learning process, as will be described in the sections that follow, we monitored model accuracy, as well as assessed the importance of different features and their gini values for identifying the species that are most relevant for differentiating habitat types. Results We observed variations in biodiversity depending on the management measures adopted at a

particular site. In total, we recorded presence of 164 vascular plant species, 31 insect pollinator species, and 61 bird species (S1 Table). At 85, 14, and 26 (G), and 115, 21, and 19 (M), the greatest number of plant, pollinator, and bird species was found at the grazed and mowed sites, respectively, while at unmanaged site (C), only 59, 9, and 18 species were recorded. The greatest differences between the sites were found in the plant species, namely the shrubs such as Cornus mas, Crataegus monogyna, Rosa canina, Prunus spinosa, and Fraxinus ornus, which dominated within the unmanaged area, while in the grazed and mown areas they appeared only sporadically and mainly as young plants. At all sites (especially those under grazing), prevalence of species from the Poaceae family was noted, and Festuca was the pre- dominant plant genus. Maximum vegetation height (MVH) ranged from 12 to 80 cm, while total vegetation cover (TVC) ranged from 10% to 100%. The G sites exhibited the significantly PLOS ONE Effects of habitat management on biodiversity PLOS ONE | https://doi.org/10.1371/journal.pone.0301391 March 28, 2024 7 / 19