| Abstract: |
Habitat protection is a critical aspect of species conservation, as restoring a habitat to its former state after it has been destroyed can be difficult. Species Distribution Models (SDMs), also known as habitat suitability models, are commonly used to address this issue. It finds ecological and evolutionary insights by linking species occurrences records to environmental data. Machine learning (ML) algorithms have been recently used to predict the distribution of species. Yet, a single ML algorithm may not always yield accurate predictions for a given dataset, making it challenging to develop a highly accurate model using a single algorithm. Therefore, this study proposes a novel approach to assess habitat suitability of three redstarts species based on ensemble learning techniques. Initially, eight machine learning algorithms, including MultiLayer Perceptron (MLP), Support Vector Machine (SVM), K-nearest neighbors (KNN), Decision Trees (DT), Gradient Boosting Classifier (GB), Random Forest (RF), AdaBoost (AB), and Quadratic Discriminant Analysis (QDA), were trained as base-learners. Subsequently, based on the performance of these base-learners, seven heterogeneous ensembles of two up to eight models, were constructed for each species dataset. The performance of the proposed approach was evaluated using five performance criteria (accuracy, sensitivity, specificity, AUC, and Kappa), Scott Knott (SK) test to statistically compare the performance of the presented models, and the Borda Count voting method to rank the best performing models based on multiple performance criteria. The findings revealed that the heterogeneous ensembles outperformed their singles in all three species datasets, underscoring the efficacy of the proposed approach in modelling species distribution. |