Improved Prediction of Eurasian Beaver Gnawing Preferences in Riparian Habitats: A Machine Learning Approach

In this study, we investigated the impact of Eurasian beavers (Castor fiber Linnaeus, 1758) on riparian woodlands in Central Italy using Machine Learning (ML) techniques. Beavers are ecosystem engineers who may modify riverine ecosystems through dam building and foraging activities. Their gnawing activity can significantly alter the composition and structure of riparian forests. Traditionally, statistical models have been used to understand factors influencing beaver activity. Thus, this study explores the potential of ML algorithms for this purpose. We implemented three ML algorithms—Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forests (RF)—to analyze data collected from three Italian rivers. Data included in situ measurements of trees (diameter, distance from riverbank, species) and information on beaver damage (i.e., signs of gnawing activity and the type of impact on the single stem). A two-step implementation has been proposed to predict whether a tree would be damaged by beavers and, if so, the severity of the damage (“low” or “high”). In the first step, three algorithms achieved high accuracy (up to 93% of damaged/undamaged trees correctly classified) and kept satisfactory performances even when trained with small subsets of the data (85% accuracy when trained with 20% of the data). In the second step, implemented over the subset of trees classified as damaged to identify those with low or high damage severity classes, the algorithms reached accuracy (85%) comparable to Step 1, despite the smaller subset available (159 samples out of 476 in the total dataset). This suggests that ML could significantly reduce the amount of field data collection needed to assess beaver impacts. The results of the analysis also suggested RF as the most suitable ML method for this kind of application in terms of both accuracy and computational cost. Moreover, the following key factors influencing beaver gnawing activity were identified: tree diameter and distance from the riverbank were the most important predictors, while tree species and site location had less influence. In summary, this study showed the potential of ML for analyzing beaver-woodland interactions with a more effective and cost-efficient sampling effort and better understanding the main factor influencing beaver gnawing activity. Future research should test the dataset from different geographical ranges, as well as incorporating data on long-term foraging sites.