Species distribution models (SDM) are key tools to obtaining knowledge about the spatial distribution of species, often achieved through ecological niche modelling, where environmental features are used to predict where species may occur.
Based on a conference challenge, this paper explores the use of deep neural networks (NN) in species distribution modelling, introducing a novel, convolutional NN model based on environmental features and co-occurrences (species interdependencies) combined, and comparing its performance to three other SDM approaches.
The authors use a GBIF-mediated dataset of 3,300 plant species occurring in France, of which three quarters were used for training models and one quarter for testing. They showed that all models performed well, with environment-based models scoring highest. While the novel NN combined with co-occurrence data did get the highest score, its performance was not significantly higher than the model based on environmental features alone.
The study highlights that while NN can improve the performance of SDM, performance of models based solely on environmental data is outperformed by models capturing species interdependencies as well.