A Machine Learning Approach for Predicting Selective Browsing Behavior

Ungulates shape forest ecosystems through the selective browsing of different plant species, which serves a regulatory role in the interspecific competition between plants. Therefore, accurately predicting the extent of browsing on different plant species may promote forest management policies which advocate biodiversity alongside economic interests. This study applied a machine learning (ML) approach to create an accurate and robust system for predicting selective browsing behavior.

Using data from seven forested areas in Hungary, three ML models were constructed to predict the extent of browsing on different plant species based on the available plant supply within the habitat: a Random Forest, Gradient Boosting, and a Zero-Inflated Beta Regressor. The models were evaluated in terms of mean absolute prediction error and the ability to correctly predict larger browsing extents. The latter is a challenging task due to the sparse and mostly small-valued nature of the data. Additionally, in light of the effort required to collect browsing data, the models’ robustness against smaller sample sizes was tested. The Gradient Boosting Regressor performed best in every aspect, achieving the highest accuracy even with lower sample sizes. A two-way ANOVA test confirmed that both the choice of ML model and the sample size has a significant effect on the prediction performance. The Gradient Boosting Regressor was integrated into a Jupyter Notebook pipeline designed for easy re-use on unseen forest data.

The results suggest that ML offers a viable approach to predict selective browsing in forest habitats, and such models can be integrated into simple applications for forest management and research applications. Future research could focus on testing the approach in a wider variety of habitats and on increasing the ease-of-use of the technique for users less familiar with ML.