Habitat Prediction and Knowledge Extraction for Marine Bivalves using Machine Learning Techniques

Species distribution models (SDMs) are powerful tools for analyzing the relationships between species and the environment. SDM results can provide insights into a species’ response to a given habitat condition, making it crucial to compare SDMs based on their predictive performance and habitat information. The marine bivalves’ habitat has been highly threatened due to anthropogenic activities and natural disturbances and continues to lose their rich biodiversity resources. Protection for these species requires detailed spatial distribution of these habitats such as habitat suitability maps. Three machine learning methods (Maximum Entropy, Random Forest, and Support Vector Machine) and Artificial Neural Network (ANN) models were used to predict the habitat suitability for marine bivalves, comparing their predictive performance and ecological relevance. A spatial modeling approach was used, incorporating 1200 occurrence data points and ten environmental factors. The study used five performance metrics to estimate the accuracy of the habitat suitability models. All of the four SDMs methods showed significant relationship between the marine bivalves distribution and environmental factors. Results indicate that Random Forest (RF) model is the best predictor of potential bivalve habitat, with an area under curve (AUC) value of 0.98 compared to SVM (0.87), MaxEnt (0.97) and ANN (0.87) models. The most important environmental factors that affect the bivalve’s distribution in the area were pH, diffuse, and calcite. Finally, a potential area for cultivating the marine bivalves with high and very suitability was suggested based on the RF model.