[Paper Review] Deep Multi-Species Embedding

Di Chen, Yexiang Xue+3 more|arXiv (Cornell University)|Sep 28, 2016

Species Distribution and Climate Change20 references5 citations

TL;DR

This paper proposes Deep Multi-Species Embedding (DMSE), a deep neural network framework that jointly embeds multiple bird species and environmental covariates into a shared high-dimensional space to model species co-occurrence and inter-species correlations. By leveraging learned representations from deep feature extraction, DMSE significantly outperforms single-species models and multi-label baselines in predicting species distributions, particularly as the number of species increases.

ABSTRACT

Understanding how species are distributed across landscapes over time is a fundamental question in biodiversity research. Unfortunately, most species distribution models only target a single species at a time, despite strong ecological evidence that species are not independently distributed. We propose Deep Multi-Species Embedding (DMSE), which jointly embeds vectors corresponding to multiple species as well as vectors representing environmental covariates into a common high-dimensional feature space via a deep neural network. Applied to bird observational data from the citizen science project extit{eBird}, we demonstrate how the DMSE model discovers inter-species relationships to outperform single-species distribution models (random forests and SVMs) as well as competing multi-label models. Additionally, we demonstrate the benefit of using a deep neural network to extract features within the embedding and show how they improve the predictive performance of species distribution modelling. An important domain contribution of the DMSE model is the ability to discover and describe species interactions while simultaneously learning the shared habitat preferences among species. As an additional contribution, we provide a graphical embedding of hundreds of bird species in the Northeast US.

Motivation & Objective

• Address the limitation of single-species species distribution models that ignore ecological interactions and co-occurrence patterns.
• Develop a scalable, joint modeling approach for hundreds of species that captures both shared environmental preferences and inter-species correlations.
• Improve predictive performance in species distribution modeling by integrating deep neural networks for feature extraction from environmental covariates.
• Enable interpretable visualization of species relationships and habitat preferences through learned vector embeddings.
• Quantify species interactions in a data-driven, scalable manner, moving beyond qualitative ecological descriptions.

Proposed method

• Use a deep neural network to map species identities and environmental covariates into a shared high-dimensional embedding space.
• Model species presence/absence using a probit link function, where the probability of detection is determined by the latent variable exceeding zero.
• Employ inner products between embedded vectors to capture semantic relationships, such as environmental preference similarity and species co-occurrence correlation.
• Train the model end-to-end using log-likelihood optimization on eBird observational data, with shared parameters across species.
• Integrate a projection matrix and deep non-linear transformations to extract hierarchical features from environmental covariates.
• Visualize embeddings using dimensionality reduction (e.g., t-SNE) to reveal spatial patterns of species co-occurrence and habitat preferences in the Northeast US.

Experimental results

Research questions

• RQ1 Can a joint embedding model effectively capture inter-species correlations in species co-occurrence while modeling environmental preferences?
• RQ2 Does incorporating deep neural networks for feature extraction improve predictive performance in multi-species distribution modeling compared to linear or shallow methods?
• RQ3 How does modeling species correlations impact predictive accuracy, especially as the number of species increases?
• RQ4 To what extent can learned embeddings reveal interpretable ecological relationships and habitat preferences?
• RQ5 Can the model quantitatively measure species interactions that were previously only qualitatively described by ecologists?

Key findings

• The DMSE model significantly outperforms single-species models (random forests and SVMs) and competing multi-label models in species distribution prediction, with performance gains increasing as the number of species grows.
• The inclusion of a deep neural network for feature extraction leads to a substantial improvement in predictive power, as shown by higher AUC scores compared to a non-deep baseline.
• Multi-species DMSE outperforms the single-species version of DMSE across all tested species pairs, demonstrating the critical role of modeling inter-species correlations.
• Species pairs with high correlation (e.g., Red-eyed Vireo and Eastern Phoebe, r = 0.607) show strong co-occurrence patterns that are captured and quantified by the model.
• The model’s performance gap widens with increasing species count, as shown in Figure 7, where DMSE consistently outperforms single-species DMSE and ensemble classifier chains.
• The graphical embedding of hundreds of bird species in the Northeast US provides an interpretable, intuitive visualization of species environmental preferences and co-occurrence patterns.