Machine Learning Prediction of Wildlife Reintroduction Success Using Ecological Indicators

Abstract

Introduction: Reintroduction of wildlife tends to experience high levels of failure because of the many interacting factors that affect post-reintroduction survival and population establishment in an unpredictable manner, therefore, the need for accurate predictive models.

Materials and Methods: In this research work, a supervised machine learning model has been used to forecast the successful reintroduction of animals based on eight environmental parameters: habitat suitability, genetic diversity, food sources, number of founders, predator abundance, climate similarity, disease presence, and human interference. Classification models that have been used are Random Forest, XGBoost, SVM, Logistic Regression, ANN, and k-NN. They have been developed after preprocessing of the data, by splitting them in 80:20 ratio through stratified sampling and hyperparameter tuning using grid search and 10-fold cross-validation.

Results: XGBoost yielded the highest level of accuracy (Accuracy = 0.91, AUC-ROC = 0.94) with a narrow margin behind it being provided by Random Forest (Accuracy = 0.89, AUC-ROC = 0.93), significantly exceeding the level of performance achieved by the linear method - Logistic Regression (Accuracy = 0.76, AUC-ROC = 0.80). This result supports the fact that non-linear relationships between features play the most crucial role in determining the success of animal reintroduction process. The feature importance evaluation has shown that habitat suitability (22.4%) and genetic diversity (18.1%) were the two leading predictors, representing more than 40% of model importance.

Conclusion: These results show that ensemble machine learning algorithms are effective interpretive tools for site prioritization and source population selection, thereby giving conservation biologists a scientific approach for the development of reintroduction strategies and efficient use of resources