Impacts of base learners selection of heterogeneous ensemble for habitat suitability modeling

Conserving species’ habitats is crucial for biodiversity conservation, as restoring them to their original state becomes challenging once they are destroyed. In this context, Species Distribution Models (SDMs), also known as Habitat Suitability Models, are widely used for this purpose. These models use environmental data and species occurrence records to predict the distribution of species and assess their habitat suitability. The application of machine learning (ML) for predicting species distribution has become increasingly popular. Nevertheless, a single ML algorithm may not provide optimal predictions for a given dataset, making it challenging to achieve high accuracy. Therefore, this study proposes a novel approach to assess habitat suitability of three bird species based on ensemble learning techniques. Initially, eight ML models were trained individually. Then, from these eight models, heterogeneous ensembles of two up to seven models were constructed for each species dataset, using two selection strategies: (1) performance-based selection and (2) diversity-based selection. The study evaluated how the diversity and performance of base-learners impact overall ensemble performance. The performance of the current experiment was evaluated using: (1) six performance measures (AUC, sensitivity, specificity, accuracy, kappa, and TSS), (2) Borda Count ranking method, (3) 95% confidence interval, and (4) Scott Knott statistical test. Heterogeneous ensembles consistently outperformed single models across all three datasets. Both performance-based and diversity-based selection strategies proved effective in improving prediction accuracy. This study showed the potential of heterogeneous ensembles for enhancing species distribution prediction, offering a new approach for improving habitat suitability assessments and supporting biodiversity conservation efforts.